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 #define MLOG_MASK_PREFIX ML_NAMEI
  47 #include <cluster/masklog.h>
  48
  49 #include "ocfs2.h"
  50
  51 #include "alloc.h"
  52 #include "blockcheck.h"
  53 #include "dir.h"
  54 #include "dlmglue.h"
  55 #include "extent_map.h"
  56 #include "file.h"
  57 #include "inode.h"
  58 #include "journal.h"
  59 #include "namei.h"
  60 #include "suballoc.h"
  61 #include "super.h"
  62 #include "sysfile.h"
  63 #include "uptodate.h"
  64
  65 #include "buffer_head_io.h"
  66
  67 #define NAMEI_RA_CHUNKS  2
  68 #define NAMEI_RA_BLOCKS  4
  69 #define NAMEI_RA_SIZE        (NAMEI_RA_CHUNKS * NAMEI_RA_BLOCKS)
  70 #define NAMEI_RA_INDEX(c,b)  (((c) * NAMEI_RA_BLOCKS) + (b))
  71
  72 static unsigned char ocfs2_filetype_table[] = {
  73         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
  74 };
  75
  76 static int ocfs2_do_extend_dir(struct super_block *sb,
  77                                handle_t *handle,
  78                                struct inode *dir,
  79                                struct buffer_head *parent_fe_bh,
  80                                struct ocfs2_alloc_context *data_ac,
  81                                struct ocfs2_alloc_context *meta_ac,
  82                                struct buffer_head **new_bh);
  83 static int ocfs2_dir_indexed(struct inode *inode);
  84
  85 /*
  86  * These are distinct checks because future versions of the file system will
  87  * want to have a trailing dirent structure independent of indexing.
  88  */
  89 static int ocfs2_supports_dir_trailer(struct inode *dir)
  90 {
  91         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
  92
  93         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
  94                 return 0;
  95
  96         return ocfs2_meta_ecc(osb) || ocfs2_dir_indexed(dir);
  97 }
  98
  99 /*
 100  * "new' here refers to the point at which we're creating a new
 101  * directory via "mkdir()", but also when we're expanding an inline
 102  * directory. In either case, we don't yet have the indexing bit set
 103  * on the directory, so the standard checks will fail in when metaecc
 104  * is turned off. Only directory-initialization type functions should
 105  * use this then. Everything else wants ocfs2_supports_dir_trailer()
 106  */
 107 static int ocfs2_new_dir_wants_trailer(struct inode *dir)
 108 {
 109         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
 110
 111         return ocfs2_meta_ecc(osb) ||
 112                 ocfs2_supports_indexed_dirs(osb);
 113 }
 114
 115 static inline unsigned int ocfs2_dir_trailer_blk_off(struct super_block *sb)
 116 {
 117         return sb->s_blocksize - sizeof(struct ocfs2_dir_block_trailer);
 118 }
 119
 120 #define ocfs2_trailer_from_bh(_bh, _sb) ((struct ocfs2_dir_block_trailer *) ((_bh)->b_data + ocfs2_dir_trailer_blk_off((_sb))))
 121
 122 /* XXX ocfs2_block_dqtrailer() is similar but not quite - can we make
 123  * them more consistent? */
 124 struct ocfs2_dir_block_trailer *ocfs2_dir_trailer_from_size(int blocksize,
 125                                                             void *data)
 126 {
 127         char *p = data;
 128
 129         p += blocksize - sizeof(struct ocfs2_dir_block_trailer);
 130         return (struct ocfs2_dir_block_trailer *)p;
 131 }
 132
 133 /*
 134  * XXX: This is executed once on every dirent. We should consider optimizing
 135  * it.
 136  */
 137 static int ocfs2_skip_dir_trailer(struct inode *dir,
 138                                   struct ocfs2_dir_entry *de,
 139                                   unsigned long offset,
 140                                   unsigned long blklen)
 141 {
 142         unsigned long toff = blklen - sizeof(struct ocfs2_dir_block_trailer);
 143
 144         if (!ocfs2_supports_dir_trailer(dir))
 145                 return 0;
 146
 147         if (offset != toff)
 148                 return 0;
 149
 150         return 1;
 151 }
 152
 153 static void ocfs2_init_dir_trailer(struct inode *inode,
 154                                    struct buffer_head *bh, u16 rec_len)
 155 {
 156         struct ocfs2_dir_block_trailer *trailer;
 157
 158         trailer = ocfs2_trailer_from_bh(bh, inode->i_sb);
 159         strcpy(trailer->db_signature, OCFS2_DIR_TRAILER_SIGNATURE);
 160         trailer->db_compat_rec_len =
 161                         cpu_to_le16(sizeof(struct ocfs2_dir_block_trailer));
 162         trailer->db_parent_dinode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
 163         trailer->db_blkno = cpu_to_le64(bh->b_blocknr);
 164         trailer->db_free_rec_len = cpu_to_le16(rec_len);
 165 }
 166 /*
 167  * Link an unindexed block with a dir trailer structure into the index free
 168  * list. This function will modify dirdata_bh, but assumes you've already
 169  * passed it to the journal.
 170  */
 171 static int ocfs2_dx_dir_link_trailer(struct inode *dir, handle_t *handle,
 172                                      struct buffer_head *dx_root_bh,
 173                                      struct buffer_head *dirdata_bh)
 174 {
 175         int ret;
 176         struct ocfs2_dx_root_block *dx_root;
 177         struct ocfs2_dir_block_trailer *trailer;
 178
 179         ret = ocfs2_journal_access_dr(handle, dir, dx_root_bh,
 180                                       OCFS2_JOURNAL_ACCESS_WRITE);
 181         if (ret) {
 182                 mlog_errno(ret);
 183                 goto out;
 184         }
 185         trailer = ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
 186         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
 187
 188         trailer->db_free_next = dx_root->dr_free_blk;
 189         dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
 190
 191         ocfs2_journal_dirty(handle, dx_root_bh);
 192
 193 out:
 194         return ret;
 195 }
 196
 197 static int ocfs2_free_list_at_root(struct ocfs2_dir_lookup_result *res)
 198 {
 199         return res->dl_prev_leaf_bh == NULL;
 200 }
 201
 202 void ocfs2_free_dir_lookup_result(struct ocfs2_dir_lookup_result *res)
 203 {
 204         brelse(res->dl_dx_root_bh);
 205         brelse(res->dl_leaf_bh);
 206         brelse(res->dl_dx_leaf_bh);
 207         brelse(res->dl_prev_leaf_bh);
 208 }
 209
 210 static int ocfs2_dir_indexed(struct inode *inode)
 211 {
 212         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INDEXED_DIR_FL)
 213                 return 1;
 214         return 0;
 215 }
 216
 217 static inline int ocfs2_dx_root_inline(struct ocfs2_dx_root_block *dx_root)
 218 {
 219         return dx_root->dr_flags & OCFS2_DX_FLAG_INLINE;
 220 }
 221
 222 /*
 223  * Hashing code adapted from ext3
 224  */
 225 #define DELTA 0x9E3779B9
 226
 227 static void TEA_transform(__u32 buf[4], __u32 const in[])
 228 {
 229         __u32   sum = 0;
 230         __u32   b0 = buf[0], b1 = buf[1];
 231         __u32   a = in[0], b = in[1], c = in[2], d = in[3];
 232         int     n = 16;
 233
 234         do {
 235                 sum += DELTA;
 236                 b0 += ((b1 << 4)+a) ^ (b1+sum) ^ ((b1 >> 5)+b);
 237                 b1 += ((b0 << 4)+c) ^ (b0+sum) ^ ((b0 >> 5)+d);
 238         } while (--n);
 239
 240         buf[0] += b0;
 241         buf[1] += b1;
 242 }
 243
 244 static void str2hashbuf(const char *msg, int len, __u32 *buf, int num)
 245 {
 246         __u32   pad, val;
 247         int     i;
 248
 249         pad = (__u32)len | ((__u32)len << 8);
 250         pad |= pad << 16;
 251
 252         val = pad;
 253         if (len > num*4)
 254                 len = num * 4;
 255         for (i = 0; i < len; i++) {
 256                 if ((i % 4) == 0)
 257                         val = pad;
 258                 val = msg[i] + (val << 8);
 259                 if ((i % 4) == 3) {
 260                         *buf++ = val;
 261                         val = pad;
 262                         num--;
 263                 }
 264         }
 265         if (--num >= 0)
 266                 *buf++ = val;
 267         while (--num >= 0)
 268                 *buf++ = pad;
 269 }
 270
 271 static void ocfs2_dx_dir_name_hash(struct inode *dir, const char *name, int len,
 272                                    struct ocfs2_dx_hinfo *hinfo)
 273 {
 274         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
 275         const char      *p;
 276         __u32           in[8], buf[4];
 277
 278         /*
 279          * XXX: Is this really necessary, if the index is never looked
 280          * at by readdir? Is a hash value of '0' a bad idea?
 281          */
 282         if ((len == 1 && !strncmp(".", name, 1)) ||
 283             (len == 2 && !strncmp("..", name, 2))) {
 284                 buf[0] = buf[1] = 0;
 285                 goto out;
 286         }
 287
 288 #ifdef OCFS2_DEBUG_DX_DIRS
 289         /*
 290          * This makes it very easy to debug indexing problems. We
 291          * should never allow this to be selected without hand editing
 292          * this file though.
 293          */
 294         buf[0] = buf[1] = len;
 295         goto out;
 296 #endif
 297
 298         memcpy(buf, osb->osb_dx_seed, sizeof(buf));
 299
 300         p = name;
 301         while (len > 0) {
 302                 str2hashbuf(p, len, in, 4);
 303                 TEA_transform(buf, in);
 304                 len -= 16;
 305                 p += 16;
 306         }
 307
 308 out:
 309         hinfo->major_hash = buf[0];
 310         hinfo->minor_hash = buf[1];
 311 }
 312
 313 /*
 314  * bh passed here can be an inode block or a dir data block, depending
 315  * on the inode inline data flag.
 316  */
 317 static int ocfs2_check_dir_entry(struct inode * dir,
 318                                  struct ocfs2_dir_entry * de,
 319                                  struct buffer_head * bh,
 320                                  unsigned long offset)
 321 {
 322         const char *error_msg = NULL;
 323         const int rlen = le16_to_cpu(de->rec_len);
 324
 325         if (rlen < OCFS2_DIR_REC_LEN(1))
 326                 error_msg = "rec_len is smaller than minimal";
 327         else if (rlen % 4 != 0)
 328                 error_msg = "rec_len % 4 != 0";
 329         else if (rlen < OCFS2_DIR_REC_LEN(de->name_len))
 330                 error_msg = "rec_len is too small for name_len";
 331         else if (((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)
 332                 error_msg = "directory entry across blocks";
 333
 334         if (error_msg != NULL)
 335                 mlog(ML_ERROR, "bad entry in directory #%llu: %s - "
 336                      "offset=%lu, inode=%llu, rec_len=%d, name_len=%d\n",
 337                      (unsigned long long)OCFS2_I(dir)->ip_blkno, error_msg,
 338                      offset, (unsigned long long)le64_to_cpu(de->inode), rlen,
 339                      de->name_len);
 340         return error_msg == NULL ? 1 : 0;
 341 }
 342
 343 static inline int ocfs2_match(int len,
 344                               const char * const name,
 345                               struct ocfs2_dir_entry *de)
 346 {
 347         if (len != de->name_len)
 348                 return 0;
 349         if (!de->inode)
 350                 return 0;
 351         return !memcmp(name, de->name, len);
 352 }
 353
 354 /*
 355  * Returns 0 if not found, -1 on failure, and 1 on success
 356  */
 357 static int inline ocfs2_search_dirblock(struct buffer_head *bh,
 358                                         struct inode *dir,
 359                                         const char *name, int namelen,
 360                                         unsigned long offset,
 361                                         char *first_de,
 362                                         unsigned int bytes,
 363                                         struct ocfs2_dir_entry **res_dir)
 364 {
 365         struct ocfs2_dir_entry *de;
 366         char *dlimit, *de_buf;
 367         int de_len;
 368         int ret = 0;
 369
 370         mlog_entry_void();
 371
 372         de_buf = first_de;
 373         dlimit = de_buf + bytes;
 374
 375         while (de_buf < dlimit) {
 376                 /* this code is executed quadratically often */
 377                 /* do minimal checking `by hand' */
 378
 379                 de = (struct ocfs2_dir_entry *) de_buf;
 380
 381                 if (de_buf + namelen <= dlimit &&
 382                     ocfs2_match(namelen, name, de)) {
 383                         /* found a match - just to be sure, do a full check */
 384                         if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
 385                                 ret = -1;
 386                                 goto bail;
 387                         }
 388                         *res_dir = de;
 389                         ret = 1;
 390                         goto bail;
 391                 }
 392
 393                 /* prevent looping on a bad block */
 394                 de_len = le16_to_cpu(de->rec_len);
 395                 if (de_len <= 0) {
 396                         ret = -1;
 397                         goto bail;
 398                 }
 399
 400                 de_buf += de_len;
 401                 offset += de_len;
 402         }
 403
 404 bail:
 405         mlog_exit(ret);
 406         return ret;
 407 }
 408
 409 static struct buffer_head *ocfs2_find_entry_id(const char *name,
 410                                                int namelen,
 411                                                struct inode *dir,
 412                                                struct ocfs2_dir_entry **res_dir)
 413 {
 414         int ret, found;
 415         struct buffer_head *di_bh = NULL;
 416         struct ocfs2_dinode *di;
 417         struct ocfs2_inline_data *data;
 418
 419         ret = ocfs2_read_inode_block(dir, &di_bh);
 420         if (ret) {
 421                 mlog_errno(ret);
 422                 goto out;
 423         }
 424
 425         di = (struct ocfs2_dinode *)di_bh->b_data;
 426         data = &di->id2.i_data;
 427
 428         found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
 429                                       data->id_data, i_size_read(dir), res_dir);
 430         if (found == 1)
 431                 return di_bh;
 432
 433         brelse(di_bh);
 434 out:
 435         return NULL;
 436 }
 437
 438 static int ocfs2_validate_dir_block(struct super_block *sb,
 439                                     struct buffer_head *bh)
 440 {
 441         int rc;
 442         struct ocfs2_dir_block_trailer *trailer =
 443                 ocfs2_trailer_from_bh(bh, sb);
 444
 445
 446         /*
 447          * We don't validate dirents here, that's handled
 448          * in-place when the code walks them.
 449          */
 450         mlog(0, "Validating dirblock %llu\n",
 451              (unsigned long long)bh->b_blocknr);
 452
 453         BUG_ON(!buffer_uptodate(bh));
 454
 455         /*
 456          * If the ecc fails, we return the error but otherwise
 457          * leave the filesystem running.  We know any error is
 458          * local to this block.
 459          *
 460          * Note that we are safe to call this even if the directory
 461          * doesn't have a trailer.  Filesystems without metaecc will do
 462          * nothing, and filesystems with it will have one.
 463          */
 464         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
 465         if (rc)
 466                 mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
 467                      (unsigned long long)bh->b_blocknr);
 468
 469         return rc;
 470 }
 471
 472 /*
 473  * Validate a directory trailer.
 474  *
 475  * We check the trailer here rather than in ocfs2_validate_dir_block()
 476  * because that function doesn't have the inode to test.
 477  */
 478 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
 479 {
 480         int rc = 0;
 481         struct ocfs2_dir_block_trailer *trailer;
 482
 483         trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
 484         if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
 485                 rc = -EINVAL;
 486                 ocfs2_error(dir->i_sb,
 487                             "Invalid dirblock #%llu: "
 488                             "signature = %.*s\n",
 489                             (unsigned long long)bh->b_blocknr, 7,
 490                             trailer->db_signature);
 491                 goto out;
 492         }
 493         if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
 494                 rc = -EINVAL;
 495                 ocfs2_error(dir->i_sb,
 496                             "Directory block #%llu has an invalid "
 497                             "db_blkno of %llu",
 498                             (unsigned long long)bh->b_blocknr,
 499                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
 500                 goto out;
 501         }
 502         if (le64_to_cpu(trailer->db_parent_dinode) !=
 503             OCFS2_I(dir)->ip_blkno) {
 504                 rc = -EINVAL;
 505                 ocfs2_error(dir->i_sb,
 506                             "Directory block #%llu on dinode "
 507                             "#%llu has an invalid parent_dinode "
 508                             "of %llu",
 509                             (unsigned long long)bh->b_blocknr,
 510                             (unsigned long long)OCFS2_I(dir)->ip_blkno,
 511                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
 512                 goto out;
 513         }
 514 out:
 515         return rc;
 516 }
 517
 518 /*
 519  * This function forces all errors to -EIO for consistency with its
 520  * predecessor, ocfs2_bread().  We haven't audited what returning the
 521  * real error codes would do to callers.  We log the real codes with
 522  * mlog_errno() before we squash them.
 523  */
 524 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
 525                                 struct buffer_head **bh, int flags)
 526 {
 527         int rc = 0;
 528         struct buffer_head *tmp = *bh;
 529
 530         rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
 531                                     ocfs2_validate_dir_block);
 532         if (rc) {
 533                 mlog_errno(rc);
 534                 goto out;
 535         }
 536
 537         if (!(flags & OCFS2_BH_READAHEAD) &&
 538             ocfs2_supports_dir_trailer(inode)) {
 539                 rc = ocfs2_check_dir_trailer(inode, tmp);
 540                 if (rc) {
 541                         if (!*bh)
 542                                 brelse(tmp);
 543                         mlog_errno(rc);
 544                         goto out;
 545                 }
 546         }
 547
 548         /* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
 549         if (!*bh)
 550                 *bh = tmp;
 551
 552 out:
 553         return rc ? -EIO : 0;
 554 }
 555
 556 /*
 557  * Read the block at 'phys' which belongs to this directory
 558  * inode. This function does no virtual->physical block translation -
 559  * what's passed in is assumed to be a valid directory block.
 560  */
 561 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
 562                                        struct buffer_head **bh)
 563 {
 564         int ret;
 565         struct buffer_head *tmp = *bh;
 566
 567         ret = ocfs2_read_block(dir, phys, &tmp, 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(dir, blkno, &tmp, ocfs2_validate_dx_root);
 626
 627         /* If ocfs2_read_block() got us a new bh, pass it up. */
 628         if (!ret && !*dx_root_bh)
 629                 *dx_root_bh = tmp;
 630
 631         return ret;
 632 }
 633
 634 static int ocfs2_validate_dx_leaf(struct super_block *sb,
 635                                   struct buffer_head *bh)
 636 {
 637         int ret;
 638         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
 639
 640         BUG_ON(!buffer_uptodate(bh));
 641
 642         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
 643         if (ret) {
 644                 mlog(ML_ERROR,
 645                      "Checksum failed for dir index leaf block %llu\n",
 646                      (unsigned long long)bh->b_blocknr);
 647                 return ret;
 648         }
 649
 650         if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
 651                 ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s",
 652                             7, dx_leaf->dl_signature);
 653                 return -EROFS;
 654         }
 655
 656         return 0;
 657 }
 658
 659 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
 660                               struct buffer_head **dx_leaf_bh)
 661 {
 662         int ret;
 663         struct buffer_head *tmp = *dx_leaf_bh;
 664
 665         ret = ocfs2_read_block(dir, blkno, &tmp, ocfs2_validate_dx_leaf);
 666
 667         /* If ocfs2_read_block() got us a new bh, pass it up. */
 668         if (!ret && !*dx_leaf_bh)
 669                 *dx_leaf_bh = tmp;
 670
 671         return ret;
 672 }
 673
 674 /*
 675  * Read a series of dx_leaf blocks. This expects all buffer_head
 676  * pointers to be NULL on function entry.
 677  */
 678 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
 679                                 struct buffer_head **dx_leaf_bhs)
 680 {
 681         int ret;
 682
 683         ret = ocfs2_read_blocks(dir, start, num, dx_leaf_bhs, 0,
 684                                 ocfs2_validate_dx_leaf);
 685         if (ret)
 686                 mlog_errno(ret);
 687
 688         return ret;
 689 }
 690
 691 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
 692                                                struct inode *dir,
 693                                                struct ocfs2_dir_entry **res_dir)
 694 {
 695         struct super_block *sb;
 696         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 697         struct buffer_head *bh, *ret = NULL;
 698         unsigned long start, block, b;
 699         int ra_max = 0;         /* Number of bh's in the readahead
 700                                    buffer, bh_use[] */
 701         int ra_ptr = 0;         /* Current index into readahead
 702                                    buffer */
 703         int num = 0;
 704         int nblocks, i, err;
 705
 706         mlog_entry_void();
 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         mlog_exit_ptr(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, el, major_hash, &eb_bh);
 806                 if (ret) {
 807                         mlog_errno(ret);
 808                         goto out;
 809                 }
 810
 811                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 812                 el = &eb->h_list;
 813
 814                 if (el->l_tree_depth) {
 815                         ocfs2_error(inode->i_sb,
 816                                     "Inode %lu has non zero tree depth in "
 817                                     "btree tree block %llu\n", inode->i_ino,
 818                                     (unsigned long long)eb_bh->b_blocknr);
 819                         ret = -EROFS;
 820                         goto out;
 821                 }
 822         }
 823
 824         found = 0;
 825         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
 826                 rec = &el->l_recs[i];
 827
 828                 if (le32_to_cpu(rec->e_cpos) <= major_hash) {
 829                         found = 1;
 830                         break;
 831                 }
 832         }
 833
 834         if (!found) {
 835                 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
 836                             "record (%u, %u, 0) in btree", inode->i_ino,
 837                             le32_to_cpu(rec->e_cpos),
 838                             ocfs2_rec_clusters(el, rec));
 839                 ret = -EROFS;
 840                 goto out;
 841         }
 842
 843         if (ret_phys_blkno)
 844                 *ret_phys_blkno = le64_to_cpu(rec->e_blkno);
 845         if (ret_cpos)
 846                 *ret_cpos = le32_to_cpu(rec->e_cpos);
 847         if (ret_clen)
 848                 *ret_clen = le16_to_cpu(rec->e_leaf_clusters);
 849
 850 out:
 851         brelse(eb_bh);
 852         return ret;
 853 }
 854
 855 /*
 856  * Returns the block index, from the start of the cluster which this
 857  * hash belongs too.
 858  */
 859 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 860                                                    u32 minor_hash)
 861 {
 862         return minor_hash & osb->osb_dx_mask;
 863 }
 864
 865 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 866                                           struct ocfs2_dx_hinfo *hinfo)
 867 {
 868         return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
 869 }
 870
 871 static int ocfs2_dx_dir_lookup(struct inode *inode,
 872                                struct ocfs2_extent_list *el,
 873                                struct ocfs2_dx_hinfo *hinfo,
 874                                u32 *ret_cpos,
 875                                u64 *ret_phys_blkno)
 876 {
 877         int ret = 0;
 878         unsigned int cend, uninitialized_var(clen);
 879         u32 uninitialized_var(cpos);
 880         u64 uninitialized_var(blkno);
 881         u32 name_hash = hinfo->major_hash;
 882
 883         ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
 884                                       &clen);
 885         if (ret) {
 886                 mlog_errno(ret);
 887                 goto out;
 888         }
 889
 890         cend = cpos + clen;
 891         if (name_hash >= cend) {
 892                 /* We want the last cluster */
 893                 blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
 894                 cpos += clen - 1;
 895         } else {
 896                 blkno += ocfs2_clusters_to_blocks(inode->i_sb,
 897                                                   name_hash - cpos);
 898                 cpos = name_hash;
 899         }
 900
 901         /*
 902          * We now have the cluster which should hold our entry. To
 903          * find the exact block from the start of the cluster to
 904          * search, we take the lower bits of the hash.
 905          */
 906         blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
 907
 908         if (ret_phys_blkno)
 909                 *ret_phys_blkno = blkno;
 910         if (ret_cpos)
 911                 *ret_cpos = cpos;
 912
 913 out:
 914
 915         return ret;
 916 }
 917
 918 static int ocfs2_dx_dir_search(const char *name, int namelen,
 919                                struct inode *dir,
 920                                struct ocfs2_dx_root_block *dx_root,
 921                                struct ocfs2_dir_lookup_result *res)
 922 {
 923         int ret, i, found;
 924         u64 uninitialized_var(phys);
 925         struct buffer_head *dx_leaf_bh = NULL;
 926         struct ocfs2_dx_leaf *dx_leaf;
 927         struct ocfs2_dx_entry *dx_entry = NULL;
 928         struct buffer_head *dir_ent_bh = NULL;
 929         struct ocfs2_dir_entry *dir_ent = NULL;
 930         struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
 931         struct ocfs2_extent_list *dr_el;
 932         struct ocfs2_dx_entry_list *entry_list;
 933
 934         ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
 935
 936         if (ocfs2_dx_root_inline(dx_root)) {
 937                 entry_list = &dx_root->dr_entries;
 938                 goto search;
 939         }
 940
 941         dr_el = &dx_root->dr_list;
 942
 943         ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
 944         if (ret) {
 945                 mlog_errno(ret);
 946                 goto out;
 947         }
 948
 949         mlog(0, "Dir %llu: name: \"%.*s\", lookup of hash: %u.0x%x "
 950              "returns: %llu\n",
 951              (unsigned long long)OCFS2_I(dir)->ip_blkno,
 952              namelen, name, hinfo->major_hash, hinfo->minor_hash,
 953              (unsigned long long)phys);
 954
 955         ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
 956         if (ret) {
 957                 mlog_errno(ret);
 958                 goto out;
 959         }
 960
 961         dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
 962
 963         mlog(0, "leaf info: num_used: %d, count: %d\n",
 964              le16_to_cpu(dx_leaf->dl_list.de_num_used),
 965              le16_to_cpu(dx_leaf->dl_list.de_count));
 966
 967         entry_list = &dx_leaf->dl_list;
 968
 969 search:
 970         /*
 971          * Empty leaf is legal, so no need to check for that.
 972          */
 973         found = 0;
 974         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
 975                 dx_entry = &entry_list->de_entries[i];
 976
 977                 if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
 978                     || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
 979                         continue;
 980
 981                 /*
 982                  * Search unindexed leaf block now. We're not
 983                  * guaranteed to find anything.
 984                  */
 985                 ret = ocfs2_read_dir_block_direct(dir,
 986                                           le64_to_cpu(dx_entry->dx_dirent_blk),
 987                                           &dir_ent_bh);
 988                 if (ret) {
 989                         mlog_errno(ret);
 990                         goto out;
 991                 }
 992
 993                 /*
 994                  * XXX: We should check the unindexed block here,
 995                  * before using it.
 996                  */
 997
 998                 found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
 999                                               0, dir_ent_bh->b_data,
1000                                               dir->i_sb->s_blocksize, &dir_ent);
1001                 if (found == 1)
1002                         break;
1003
1004                 if (found == -1) {
1005                         /* This means we found a bad directory entry. */
1006                         ret = -EIO;
1007                         mlog_errno(ret);
1008                         goto out;
1009                 }
1010
1011                 brelse(dir_ent_bh);
1012                 dir_ent_bh = NULL;
1013         }
1014
1015         if (found <= 0) {
1016                 ret = -ENOENT;
1017                 goto out;
1018         }
1019
1020         res->dl_leaf_bh = dir_ent_bh;
1021         res->dl_entry = dir_ent;
1022         res->dl_dx_leaf_bh = dx_leaf_bh;
1023         res->dl_dx_entry = dx_entry;
1024
1025         ret = 0;
1026 out:
1027         if (ret) {
1028                 brelse(dx_leaf_bh);
1029                 brelse(dir_ent_bh);
1030         }
1031         return ret;
1032 }
1033
1034 static int ocfs2_find_entry_dx(const char *name, int namelen,
1035                                struct inode *dir,
1036                                struct ocfs2_dir_lookup_result *lookup)
1037 {
1038         int ret;
1039         struct buffer_head *di_bh = NULL;
1040         struct ocfs2_dinode *di;
1041         struct buffer_head *dx_root_bh = NULL;
1042         struct ocfs2_dx_root_block *dx_root;
1043
1044         ret = ocfs2_read_inode_block(dir, &di_bh);
1045         if (ret) {
1046                 mlog_errno(ret);
1047                 goto out;
1048         }
1049
1050         di = (struct ocfs2_dinode *)di_bh->b_data;
1051
1052         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1053         if (ret) {
1054                 mlog_errno(ret);
1055                 goto out;
1056         }
1057         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1058
1059         ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1060         if (ret) {
1061                 if (ret != -ENOENT)
1062                         mlog_errno(ret);
1063                 goto out;
1064         }
1065
1066         lookup->dl_dx_root_bh = dx_root_bh;
1067         dx_root_bh = NULL;
1068 out:
1069         brelse(di_bh);
1070         brelse(dx_root_bh);
1071         return ret;
1072 }
1073
1074 /*
1075  * Try to find an entry of the provided name within 'dir'.
1076  *
1077  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1078  * returned and the struct 'res' will contain information useful to
1079  * other directory manipulation functions.
1080  *
1081  * Caller can NOT assume anything about the contents of the
1082  * buffer_heads - they are passed back only so that it can be passed
1083  * into any one of the manipulation functions (add entry, delete
1084  * entry, etc). As an example, bh in the extent directory case is a
1085  * data block, in the inline-data case it actually points to an inode,
1086  * in the indexed directory case, multiple buffers are involved.
1087  */
1088 int ocfs2_find_entry(const char *name, int namelen,
1089                      struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1090 {
1091         struct buffer_head *bh;
1092         struct ocfs2_dir_entry *res_dir = NULL;
1093
1094         if (ocfs2_dir_indexed(dir))
1095                 return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1096
1097         /*
1098          * The unindexed dir code only uses part of the lookup
1099          * structure, so there's no reason to push it down further
1100          * than this.
1101          */
1102         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1103                 bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1104         else
1105                 bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1106
1107         if (bh == NULL)
1108                 return -ENOENT;
1109
1110         lookup->dl_leaf_bh = bh;
1111         lookup->dl_entry = res_dir;
1112         return 0;
1113 }
1114
1115 /*
1116  * Update inode number and type of a previously found directory entry.
1117  */
1118 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1119                        struct ocfs2_dir_lookup_result *res,
1120                        struct inode *new_entry_inode)
1121 {
1122         int ret;
1123         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1124         struct ocfs2_dir_entry *de = res->dl_entry;
1125         struct buffer_head *de_bh = res->dl_leaf_bh;
1126
1127         /*
1128          * The same code works fine for both inline-data and extent
1129          * based directories, so no need to split this up.  The only
1130          * difference is the journal_access function.
1131          */
1132
1133         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1134                 access = ocfs2_journal_access_di;
1135
1136         ret = access(handle, dir, de_bh, 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         mlog_entry("(0x%p, 0x%p, 0x%p, 0x%p)\n", handle, dir, de_del, bh);
1165
1166         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1167                 access = ocfs2_journal_access_di;
1168
1169         i = 0;
1170         pde = NULL;
1171         de = (struct ocfs2_dir_entry *) first_de;
1172         while (i < bytes) {
1173                 if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1174                         status = -EIO;
1175                         mlog_errno(status);
1176                         goto bail;
1177                 }
1178                 if (de == de_del)  {
1179                         status = access(handle, dir, bh,
1180                                         OCFS2_JOURNAL_ACCESS_WRITE);
1181                         if (status < 0) {
1182                                 status = -EIO;
1183                                 mlog_errno(status);
1184                                 goto bail;
1185                         }
1186                         if (pde)
1187                                 le16_add_cpu(&pde->rec_len,
1188                                                 le16_to_cpu(de->rec_len));
1189                         else
1190                                 de->inode = 0;
1191                         dir->i_version++;
1192                         status = ocfs2_journal_dirty(handle, bh);
1193                         goto bail;
1194                 }
1195                 i += le16_to_cpu(de->rec_len);
1196                 pde = de;
1197                 de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1198         }
1199 bail:
1200         mlog_exit(status);
1201         return status;
1202 }
1203
1204 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1205 {
1206         unsigned int hole;
1207
1208         if (le64_to_cpu(de->inode) == 0)
1209                 hole = le16_to_cpu(de->rec_len);
1210         else
1211                 hole = le16_to_cpu(de->rec_len) -
1212                         OCFS2_DIR_REC_LEN(de->name_len);
1213
1214         return hole;
1215 }
1216
1217 static int ocfs2_find_max_rec_len(struct super_block *sb,
1218                                   struct buffer_head *dirblock_bh)
1219 {
1220         int size, this_hole, largest_hole = 0;
1221         char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1222         struct ocfs2_dir_entry *de;
1223
1224         trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1225         size = ocfs2_dir_trailer_blk_off(sb);
1226         limit = start + size;
1227         de_buf = start;
1228         de = (struct ocfs2_dir_entry *)de_buf;
1229         do {
1230                 if (de_buf != trailer) {
1231                         this_hole = ocfs2_figure_dirent_hole(de);
1232                         if (this_hole > largest_hole)
1233                                 largest_hole = this_hole;
1234                 }
1235
1236                 de_buf += le16_to_cpu(de->rec_len);
1237                 de = (struct ocfs2_dir_entry *)de_buf;
1238         } while (de_buf < limit);
1239
1240         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1241                 return largest_hole;
1242         return 0;
1243 }
1244
1245 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1246                                        int index)
1247 {
1248         int num_used = le16_to_cpu(entry_list->de_num_used);
1249
1250         if (num_used == 1 || index == (num_used - 1))
1251                 goto clear;
1252
1253         memmove(&entry_list->de_entries[index],
1254                 &entry_list->de_entries[index + 1],
1255                 (num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1256 clear:
1257         num_used--;
1258         memset(&entry_list->de_entries[num_used], 0,
1259                sizeof(struct ocfs2_dx_entry));
1260         entry_list->de_num_used = cpu_to_le16(num_used);
1261 }
1262
1263 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1264                                  struct ocfs2_dir_lookup_result *lookup)
1265 {
1266         int ret, index, max_rec_len, add_to_free_list = 0;
1267         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1268         struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1269         struct ocfs2_dx_leaf *dx_leaf;
1270         struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1271         struct ocfs2_dir_block_trailer *trailer;
1272         struct ocfs2_dx_root_block *dx_root;
1273         struct ocfs2_dx_entry_list *entry_list;
1274
1275         /*
1276          * This function gets a bit messy because we might have to
1277          * modify the root block, regardless of whether the indexed
1278          * entries are stored inline.
1279          */
1280
1281         /*
1282          * *Only* set 'entry_list' here, based on where we're looking
1283          * for the indexed entries. Later, we might still want to
1284          * journal both blocks, based on free list state.
1285          */
1286         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1287         if (ocfs2_dx_root_inline(dx_root)) {
1288                 entry_list = &dx_root->dr_entries;
1289         } else {
1290                 dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1291                 entry_list = &dx_leaf->dl_list;
1292         }
1293
1294         /* Neither of these are a disk corruption - that should have
1295          * been caught by lookup, before we got here. */
1296         BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1297         BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1298
1299         index = (char *)dx_entry - (char *)entry_list->de_entries;
1300         index /= sizeof(*dx_entry);
1301
1302         if (index >= le16_to_cpu(entry_list->de_num_used)) {
1303                 mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1304                      (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1305                      entry_list, dx_entry);
1306                 return -EIO;
1307         }
1308
1309         /*
1310          * We know that removal of this dirent will leave enough room
1311          * for a new one, so add this block to the free list if it
1312          * isn't already there.
1313          */
1314         trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1315         if (trailer->db_free_rec_len == 0)
1316                 add_to_free_list = 1;
1317
1318         /*
1319          * Add the block holding our index into the journal before
1320          * removing the unindexed entry. If we get an error return
1321          * from __ocfs2_delete_entry(), then it hasn't removed the
1322          * entry yet. Likewise, successful return means we *must*
1323          * remove the indexed entry.
1324          *
1325          * We're also careful to journal the root tree block here as
1326          * the entry count needs to be updated. Also, we might be
1327          * adding to the start of the free list.
1328          */
1329         ret = ocfs2_journal_access_dr(handle, dir, dx_root_bh,
1330                                       OCFS2_JOURNAL_ACCESS_WRITE);
1331         if (ret) {
1332                 mlog_errno(ret);
1333                 goto out;
1334         }
1335
1336         if (!ocfs2_dx_root_inline(dx_root)) {
1337                 ret = ocfs2_journal_access_dl(handle, dir,
1338                                               lookup->dl_dx_leaf_bh,
1339                                               OCFS2_JOURNAL_ACCESS_WRITE);
1340                 if (ret) {
1341                         mlog_errno(ret);
1342                         goto out;
1343                 }
1344         }
1345
1346         mlog(0, "Dir %llu: delete entry at index: %d\n",
1347              (unsigned long long)OCFS2_I(dir)->ip_blkno, index);
1348
1349         ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1350                                    leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1351         if (ret) {
1352                 mlog_errno(ret);
1353                 goto out;
1354         }
1355
1356         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1357         trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1358         if (add_to_free_list) {
1359                 trailer->db_free_next = dx_root->dr_free_blk;
1360                 dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1361                 ocfs2_journal_dirty(handle, dx_root_bh);
1362         }
1363
1364         /* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1365         ocfs2_journal_dirty(handle, leaf_bh);
1366
1367         le32_add_cpu(&dx_root->dr_num_entries, -1);
1368         ocfs2_journal_dirty(handle, dx_root_bh);
1369
1370         ocfs2_dx_list_remove_entry(entry_list, index);
1371
1372         if (!ocfs2_dx_root_inline(dx_root))
1373                 ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1374
1375 out:
1376         return ret;
1377 }
1378
1379 static inline int ocfs2_delete_entry_id(handle_t *handle,
1380                                         struct inode *dir,
1381                                         struct ocfs2_dir_entry *de_del,
1382                                         struct buffer_head *bh)
1383 {
1384         int ret;
1385         struct buffer_head *di_bh = NULL;
1386         struct ocfs2_dinode *di;
1387         struct ocfs2_inline_data *data;
1388
1389         ret = ocfs2_read_inode_block(dir, &di_bh);
1390         if (ret) {
1391                 mlog_errno(ret);
1392                 goto out;
1393         }
1394
1395         di = (struct ocfs2_dinode *)di_bh->b_data;
1396         data = &di->id2.i_data;
1397
1398         ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1399                                    i_size_read(dir));
1400
1401         brelse(di_bh);
1402 out:
1403         return ret;
1404 }
1405
1406 static inline int ocfs2_delete_entry_el(handle_t *handle,
1407                                         struct inode *dir,
1408                                         struct ocfs2_dir_entry *de_del,
1409                                         struct buffer_head *bh)
1410 {
1411         return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1412                                     bh->b_size);
1413 }
1414
1415 /*
1416  * Delete a directory entry. Hide the details of directory
1417  * implementation from the caller.
1418  */
1419 int ocfs2_delete_entry(handle_t *handle,
1420                        struct inode *dir,
1421                        struct ocfs2_dir_lookup_result *res)
1422 {
1423         if (ocfs2_dir_indexed(dir))
1424                 return ocfs2_delete_entry_dx(handle, dir, res);
1425
1426         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1427                 return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1428                                              res->dl_leaf_bh);
1429
1430         return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1431                                      res->dl_leaf_bh);
1432 }
1433
1434 /*
1435  * Check whether 'de' has enough room to hold an entry of
1436  * 'new_rec_len' bytes.
1437  */
1438 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1439                                          unsigned int new_rec_len)
1440 {
1441         unsigned int de_really_used;
1442
1443         /* Check whether this is an empty record with enough space */
1444         if (le64_to_cpu(de->inode) == 0 &&
1445             le16_to_cpu(de->rec_len) >= new_rec_len)
1446                 return 1;
1447
1448         /*
1449          * Record might have free space at the end which we can
1450          * use.
1451          */
1452         de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1453         if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1454             return 1;
1455
1456         return 0;
1457 }
1458
1459 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1460                                           struct ocfs2_dx_entry *dx_new_entry)
1461 {
1462         int i;
1463
1464         i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1465         dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1466
1467         le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1468 }
1469
1470 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1471                                        struct ocfs2_dx_hinfo *hinfo,
1472                                        u64 dirent_blk)
1473 {
1474         int i;
1475         struct ocfs2_dx_entry *dx_entry;
1476
1477         i = le16_to_cpu(entry_list->de_num_used);
1478         dx_entry = &entry_list->de_entries[i];
1479
1480         memset(dx_entry, 0, sizeof(*dx_entry));
1481         dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1482         dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1483         dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1484
1485         le16_add_cpu(&entry_list->de_num_used, 1);
1486 }
1487
1488 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1489                                       struct ocfs2_dx_hinfo *hinfo,
1490                                       u64 dirent_blk,
1491                                       struct buffer_head *dx_leaf_bh)
1492 {
1493         int ret;
1494         struct ocfs2_dx_leaf *dx_leaf;
1495
1496         ret = ocfs2_journal_access_dl(handle, dir, dx_leaf_bh,
1497                                       OCFS2_JOURNAL_ACCESS_WRITE);
1498         if (ret) {
1499                 mlog_errno(ret);
1500                 goto out;
1501         }
1502
1503         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1504         ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1505         ocfs2_journal_dirty(handle, dx_leaf_bh);
1506
1507 out:
1508         return ret;
1509 }
1510
1511 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1512                                         struct ocfs2_dx_hinfo *hinfo,
1513                                         u64 dirent_blk,
1514                                         struct ocfs2_dx_root_block *dx_root)
1515 {
1516         ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1517 }
1518
1519 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1520                                struct ocfs2_dir_lookup_result *lookup)
1521 {
1522         int ret = 0;
1523         struct ocfs2_dx_root_block *dx_root;
1524         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1525
1526         ret = ocfs2_journal_access_dr(handle, dir, dx_root_bh,
1527                                       OCFS2_JOURNAL_ACCESS_WRITE);
1528         if (ret) {
1529                 mlog_errno(ret);
1530                 goto out;
1531         }
1532
1533         dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1534         if (ocfs2_dx_root_inline(dx_root)) {
1535                 ocfs2_dx_inline_root_insert(dir, handle,
1536                                             &lookup->dl_hinfo,
1537                                             lookup->dl_leaf_bh->b_blocknr,
1538                                             dx_root);
1539         } else {
1540                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1541                                                  lookup->dl_leaf_bh->b_blocknr,
1542                                                  lookup->dl_dx_leaf_bh);
1543                 if (ret)
1544                         goto out;
1545         }
1546
1547         le32_add_cpu(&dx_root->dr_num_entries, 1);
1548         ocfs2_journal_dirty(handle, dx_root_bh);
1549
1550 out:
1551         return ret;
1552 }
1553
1554 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1555                                        handle_t *handle,
1556                                        struct ocfs2_dir_lookup_result *lookup)
1557 {
1558         struct ocfs2_dir_block_trailer *trailer, *prev;
1559         struct ocfs2_dx_root_block *dx_root;
1560         struct buffer_head *bh;
1561
1562         trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1563
1564         if (ocfs2_free_list_at_root(lookup)) {
1565                 bh = lookup->dl_dx_root_bh;
1566                 dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1567                 dx_root->dr_free_blk = trailer->db_free_next;
1568         } else {
1569                 bh = lookup->dl_prev_leaf_bh;
1570                 prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1571                 prev->db_free_next = trailer->db_free_next;
1572         }
1573
1574         trailer->db_free_rec_len = cpu_to_le16(0);
1575         trailer->db_free_next = cpu_to_le64(0);
1576
1577         ocfs2_journal_dirty(handle, bh);
1578         ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1579 }
1580
1581 /*
1582  * This expects that a journal write has been reserved on
1583  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1584  */
1585 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1586                                    struct ocfs2_dir_lookup_result *lookup)
1587 {
1588         int max_rec_len;
1589         struct ocfs2_dir_block_trailer *trailer;
1590
1591         /* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1592         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1593         if (max_rec_len) {
1594                 /*
1595                  * There's still room in this block, so no need to remove it
1596                  * from the free list. In this case, we just want to update
1597                  * the rec len accounting.
1598                  */
1599                 trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1600                 trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1601                 ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1602         } else {
1603                 ocfs2_remove_block_from_free_list(dir, handle, lookup);
1604         }
1605 }
1606
1607 /* we don't always have a dentry for what we want to add, so people
1608  * like orphan dir can call this instead.
1609  *
1610  * The lookup context must have been filled from
1611  * ocfs2_prepare_dir_for_insert.
1612  */
1613 int __ocfs2_add_entry(handle_t *handle,
1614                       struct inode *dir,
1615                       const char *name, int namelen,
1616                       struct inode *inode, u64 blkno,
1617                       struct buffer_head *parent_fe_bh,
1618                       struct ocfs2_dir_lookup_result *lookup)
1619 {
1620         unsigned long offset;
1621         unsigned short rec_len;
1622         struct ocfs2_dir_entry *de, *de1;
1623         struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1624         struct super_block *sb = dir->i_sb;
1625         int retval, status;
1626         unsigned int size = sb->s_blocksize;
1627         struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1628         char *data_start = insert_bh->b_data;
1629
1630         mlog_entry_void();
1631
1632         if (!namelen)
1633                 return -EINVAL;
1634
1635         if (ocfs2_dir_indexed(dir)) {
1636                 struct buffer_head *bh;
1637
1638                 /*
1639                  * An indexed dir may require that we update the free space
1640                  * list. Reserve a write to the previous node in the list so
1641                  * that we don't fail later.
1642                  *
1643                  * XXX: This can be either a dx_root_block, or an unindexed
1644                  * directory tree leaf block.
1645                  */
1646                 if (ocfs2_free_list_at_root(lookup)) {
1647                         bh = lookup->dl_dx_root_bh;
1648                         retval = ocfs2_journal_access_dr(handle, dir, bh,
1649                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1650                 } else {
1651                         bh = lookup->dl_prev_leaf_bh;
1652                         retval = ocfs2_journal_access_db(handle, dir, bh,
1653                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1654                 }
1655                 if (retval) {
1656                         mlog_errno(retval);
1657                         return retval;
1658                 }
1659         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1660                 data_start = di->id2.i_data.id_data;
1661                 size = i_size_read(dir);
1662
1663                 BUG_ON(insert_bh != parent_fe_bh);
1664         }
1665
1666         rec_len = OCFS2_DIR_REC_LEN(namelen);
1667         offset = 0;
1668         de = (struct ocfs2_dir_entry *) data_start;
1669         while (1) {
1670                 BUG_ON((char *)de >= (size + data_start));
1671
1672                 /* These checks should've already been passed by the
1673                  * prepare function, but I guess we can leave them
1674                  * here anyway. */
1675                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1676                         retval = -ENOENT;
1677                         goto bail;
1678                 }
1679                 if (ocfs2_match(namelen, name, de)) {
1680                         retval = -EEXIST;
1681                         goto bail;
1682                 }
1683
1684                 /* We're guaranteed that we should have space, so we
1685                  * can't possibly have hit the trailer...right? */
1686                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1687                                 "Hit dir trailer trying to insert %.*s "
1688                                 "(namelen %d) into directory %llu.  "
1689                                 "offset is %lu, trailer offset is %d\n",
1690                                 namelen, name, namelen,
1691                                 (unsigned long long)parent_fe_bh->b_blocknr,
1692                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1693
1694                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1695                         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
1696                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1697                         if (retval < 0) {
1698                                 mlog_errno(retval);
1699                                 goto bail;
1700                         }
1701
1702                         if (insert_bh == parent_fe_bh)
1703                                 status = ocfs2_journal_access_di(handle, dir,
1704                                                                  insert_bh,
1705                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1706                         else {
1707                                 status = ocfs2_journal_access_db(handle, dir,
1708                                                                  insert_bh,
1709                                               OCFS2_JOURNAL_ACCESS_WRITE);
1710
1711                                 if (ocfs2_dir_indexed(dir)) {
1712                                         status = ocfs2_dx_dir_insert(dir,
1713                                                                 handle,
1714                                                                 lookup);
1715                                         if (status) {
1716                                                 mlog_errno(status);
1717                                                 goto bail;
1718                                         }
1719                                 }
1720                         }
1721
1722                         /* By now the buffer is marked for journaling */
1723                         offset += le16_to_cpu(de->rec_len);
1724                         if (le64_to_cpu(de->inode)) {
1725                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1726                                         OCFS2_DIR_REC_LEN(de->name_len));
1727                                 de1->rec_len =
1728                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1729                                         OCFS2_DIR_REC_LEN(de->name_len));
1730                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1731                                 de = de1;
1732                         }
1733                         de->file_type = OCFS2_FT_UNKNOWN;
1734                         if (blkno) {
1735                                 de->inode = cpu_to_le64(blkno);
1736                                 ocfs2_set_de_type(de, inode->i_mode);
1737                         } else
1738                                 de->inode = 0;
1739                         de->name_len = namelen;
1740                         memcpy(de->name, name, namelen);
1741
1742                         if (ocfs2_dir_indexed(dir))
1743                                 ocfs2_recalc_free_list(dir, handle, lookup);
1744
1745                         dir->i_version++;
1746                         status = ocfs2_journal_dirty(handle, insert_bh);
1747                         retval = 0;
1748                         goto bail;
1749                 }
1750
1751                 offset += le16_to_cpu(de->rec_len);
1752                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1753         }
1754
1755         /* when you think about it, the assert above should prevent us
1756          * from ever getting here. */
1757         retval = -ENOSPC;
1758 bail:
1759
1760         mlog_exit(retval);
1761         return retval;
1762 }
1763
1764 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1765                                     u64 *f_version,
1766                                     loff_t *f_pos, void *priv,
1767                                     filldir_t filldir, int *filldir_err)
1768 {
1769         int ret, i, filldir_ret;
1770         unsigned long offset = *f_pos;
1771         struct buffer_head *di_bh = NULL;
1772         struct ocfs2_dinode *di;
1773         struct ocfs2_inline_data *data;
1774         struct ocfs2_dir_entry *de;
1775
1776         ret = ocfs2_read_inode_block(inode, &di_bh);
1777         if (ret) {
1778                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1779                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1780                 goto out;
1781         }
1782
1783         di = (struct ocfs2_dinode *)di_bh->b_data;
1784         data = &di->id2.i_data;
1785
1786         while (*f_pos < i_size_read(inode)) {
1787 revalidate:
1788                 /* If the dir block has changed since the last call to
1789                  * readdir(2), then we might be pointing to an invalid
1790                  * dirent right now.  Scan from the start of the block
1791                  * to make sure. */
1792                 if (*f_version != inode->i_version) {
1793                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1794                                 de = (struct ocfs2_dir_entry *)
1795                                         (data->id_data + i);
1796                                 /* It's too expensive to do a full
1797                                  * dirent test each time round this
1798                                  * loop, but we do have to test at
1799                                  * least that it is non-zero.  A
1800                                  * failure will be detected in the
1801                                  * dirent test below. */
1802                                 if (le16_to_cpu(de->rec_len) <
1803                                     OCFS2_DIR_REC_LEN(1))
1804                                         break;
1805                                 i += le16_to_cpu(de->rec_len);
1806                         }
1807                         *f_pos = offset = i;
1808                         *f_version = inode->i_version;
1809                 }
1810
1811                 de = (struct ocfs2_dir_entry *) (data->id_data + *f_pos);
1812                 if (!ocfs2_check_dir_entry(inode, de, di_bh, *f_pos)) {
1813                         /* On error, skip the f_pos to the end. */
1814                         *f_pos = i_size_read(inode);
1815                         goto out;
1816                 }
1817                 offset += le16_to_cpu(de->rec_len);
1818                 if (le64_to_cpu(de->inode)) {
1819                         /* We might block in the next section
1820                          * if the data destination is
1821                          * currently swapped out.  So, use a
1822                          * version stamp to detect whether or
1823                          * not the directory has been modified
1824                          * during the copy operation.
1825                          */
1826                         u64 version = *f_version;
1827                         unsigned char d_type = DT_UNKNOWN;
1828
1829                         if (de->file_type < OCFS2_FT_MAX)
1830                                 d_type = ocfs2_filetype_table[de->file_type];
1831
1832                         filldir_ret = filldir(priv, de->name,
1833                                               de->name_len,
1834                                               *f_pos,
1835                                               le64_to_cpu(de->inode),
1836                                               d_type);
1837                         if (filldir_ret) {
1838                                 if (filldir_err)
1839                                         *filldir_err = filldir_ret;
1840                                 break;
1841                         }
1842                         if (version != *f_version)
1843                                 goto revalidate;
1844                 }
1845                 *f_pos += le16_to_cpu(de->rec_len);
1846         }
1847
1848 out:
1849         brelse(di_bh);
1850
1851         return 0;
1852 }
1853
1854 /*
1855  * NOTE: This function can be called against unindexed directories,
1856  * and indexed ones.
1857  */
1858 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1859                                     u64 *f_version,
1860                                     loff_t *f_pos, void *priv,
1861                                     filldir_t filldir, int *filldir_err)
1862 {
1863         int error = 0;
1864         unsigned long offset, blk, last_ra_blk = 0;
1865         int i, stored;
1866         struct buffer_head * bh, * tmp;
1867         struct ocfs2_dir_entry * de;
1868         struct super_block * sb = inode->i_sb;
1869         unsigned int ra_sectors = 16;
1870
1871         stored = 0;
1872         bh = NULL;
1873
1874         offset = (*f_pos) & (sb->s_blocksize - 1);
1875
1876         while (!error && !stored && *f_pos < i_size_read(inode)) {
1877                 blk = (*f_pos) >> sb->s_blocksize_bits;
1878                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1879                         /* Skip the corrupt dirblock and keep trying */
1880                         *f_pos += sb->s_blocksize - offset;
1881                         continue;
1882                 }
1883
1884                 /* The idea here is to begin with 8k read-ahead and to stay
1885                  * 4k ahead of our current position.
1886                  *
1887                  * TODO: Use the pagecache for this. We just need to
1888                  * make sure it's cluster-safe... */
1889                 if (!last_ra_blk
1890                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1891                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1892                              i > 0; i--) {
1893                                 tmp = NULL;
1894                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1895                                                           OCFS2_BH_READAHEAD))
1896                                         brelse(tmp);
1897                         }
1898                         last_ra_blk = blk;
1899                         ra_sectors = 8;
1900                 }
1901
1902 revalidate:
1903                 /* If the dir block has changed since the last call to
1904                  * readdir(2), then we might be pointing to an invalid
1905                  * dirent right now.  Scan from the start of the block
1906                  * to make sure. */
1907                 if (*f_version != inode->i_version) {
1908                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1909                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1910                                 /* It's too expensive to do a full
1911                                  * dirent test each time round this
1912                                  * loop, but we do have to test at
1913                                  * least that it is non-zero.  A
1914                                  * failure will be detected in the
1915                                  * dirent test below. */
1916                                 if (le16_to_cpu(de->rec_len) <
1917                                     OCFS2_DIR_REC_LEN(1))
1918                                         break;
1919                                 i += le16_to_cpu(de->rec_len);
1920                         }
1921                         offset = i;
1922                         *f_pos = ((*f_pos) & ~(sb->s_blocksize - 1))
1923                                 | offset;
1924                         *f_version = inode->i_version;
1925                 }
1926
1927                 while (!error && *f_pos < i_size_read(inode)
1928                        && offset < sb->s_blocksize) {
1929                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1930                         if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1931                                 /* On error, skip the f_pos to the
1932                                    next block. */
1933                                 *f_pos = ((*f_pos) | (sb->s_blocksize - 1)) + 1;
1934                                 brelse(bh);
1935                                 goto out;
1936                         }
1937                         offset += le16_to_cpu(de->rec_len);
1938                         if (le64_to_cpu(de->inode)) {
1939                                 /* We might block in the next section
1940                                  * if the data destination is
1941                                  * currently swapped out.  So, use a
1942                                  * version stamp to detect whether or
1943                                  * not the directory has been modified
1944                                  * during the copy operation.
1945                                  */
1946                                 unsigned long version = *f_version;
1947                                 unsigned char d_type = DT_UNKNOWN;
1948
1949                                 if (de->file_type < OCFS2_FT_MAX)
1950                                         d_type = ocfs2_filetype_table[de->file_type];
1951                                 error = filldir(priv, de->name,
1952                                                 de->name_len,
1953                                                 *f_pos,
1954                                                 le64_to_cpu(de->inode),
1955                                                 d_type);
1956                                 if (error) {
1957                                         if (filldir_err)
1958                                                 *filldir_err = error;
1959                                         break;
1960                                 }
1961                                 if (version != *f_version)
1962                                         goto revalidate;
1963                                 stored ++;
1964                         }
1965                         *f_pos += le16_to_cpu(de->rec_len);
1966                 }
1967                 offset = 0;
1968                 brelse(bh);
1969                 bh = NULL;
1970         }
1971
1972         stored = 0;
1973 out:
1974         return stored;
1975 }
1976
1977 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1978                                  loff_t *f_pos, void *priv, filldir_t filldir,
1979                                  int *filldir_err)
1980 {
1981         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1982                 return ocfs2_dir_foreach_blk_id(inode, f_version, f_pos, priv,
1983                                                 filldir, filldir_err);
1984
1985         return ocfs2_dir_foreach_blk_el(inode, f_version, f_pos, priv, filldir,
1986                                         filldir_err);
1987 }
1988
1989 /*
1990  * This is intended to be called from inside other kernel functions,
1991  * so we fake some arguments.
1992  */
1993 int ocfs2_dir_foreach(struct inode *inode, loff_t *f_pos, void *priv,
1994                       filldir_t filldir)
1995 {
1996         int ret = 0, filldir_err = 0;
1997         u64 version = inode->i_version;
1998
1999         while (*f_pos < i_size_read(inode)) {
2000                 ret = ocfs2_dir_foreach_blk(inode, &version, f_pos, priv,
2001                                             filldir, &filldir_err);
2002                 if (ret || filldir_err)
2003                         break;
2004         }
2005
2006         if (ret > 0)
2007                 ret = -EIO;
2008
2009         return 0;
2010 }
2011
2012 /*
2013  * ocfs2_readdir()
2014  *
2015  */
2016 int ocfs2_readdir(struct file * filp, void * dirent, filldir_t filldir)
2017 {
2018         int error = 0;
2019         struct inode *inode = filp->f_path.dentry->d_inode;
2020         int lock_level = 0;
2021
2022         mlog_entry("dirino=%llu\n",
2023                    (unsigned long long)OCFS2_I(inode)->ip_blkno);
2024
2025         error = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2026         if (lock_level && error >= 0) {
2027                 /* We release EX lock which used to update atime
2028                  * and get PR lock again to reduce contention
2029                  * on commonly accessed directories. */
2030                 ocfs2_inode_unlock(inode, 1);
2031                 lock_level = 0;
2032                 error = ocfs2_inode_lock(inode, NULL, 0);
2033         }
2034         if (error < 0) {
2035                 if (error != -ENOENT)
2036                         mlog_errno(error);
2037                 /* we haven't got any yet, so propagate the error. */
2038                 goto bail_nolock;
2039         }
2040
2041         error = ocfs2_dir_foreach_blk(inode, &filp->f_version, &filp->f_pos,
2042                                       dirent, filldir, NULL);
2043
2044         ocfs2_inode_unlock(inode, lock_level);
2045
2046 bail_nolock:
2047         mlog_exit(error);
2048
2049         return error;
2050 }
2051
2052 /*
2053  * NOTE: this should always be called with parent dir i_mutex taken.
2054  */
2055 int ocfs2_find_files_on_disk(const char *name,
2056                              int namelen,
2057                              u64 *blkno,
2058                              struct inode *inode,
2059                              struct ocfs2_dir_lookup_result *lookup)
2060 {
2061         int status = -ENOENT;
2062
2063         mlog(0, "name=%.*s, blkno=%p, inode=%llu\n", namelen, name, blkno,
2064              (unsigned long long)OCFS2_I(inode)->ip_blkno);
2065
2066         status = ocfs2_find_entry(name, namelen, inode, lookup);
2067         if (status)
2068                 goto leave;
2069
2070         *blkno = le64_to_cpu(lookup->dl_entry->inode);
2071
2072         status = 0;
2073 leave:
2074
2075         return status;
2076 }
2077
2078 /*
2079  * Convenience function for callers which just want the block number
2080  * mapped to a name and don't require the full dirent info, etc.
2081  */
2082 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
2083                                int namelen, u64 *blkno)
2084 {
2085         int ret;
2086         struct ocfs2_dir_lookup_result lookup = { NULL, };
2087
2088         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2089         ocfs2_free_dir_lookup_result(&lookup);
2090
2091         return ret;
2092 }
2093
2094 /* Check for a name within a directory.
2095  *
2096  * Return 0 if the name does not exist
2097  * Return -EEXIST if the directory contains the name
2098  *
2099  * Callers should have i_mutex + a cluster lock on dir
2100  */
2101 int ocfs2_check_dir_for_entry(struct inode *dir,
2102                               const char *name,
2103                               int namelen)
2104 {
2105         int ret;
2106         struct ocfs2_dir_lookup_result lookup = { NULL, };
2107
2108         mlog_entry("dir %llu, name '%.*s'\n",
2109                    (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2110
2111         ret = -EEXIST;
2112         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0)
2113                 goto bail;
2114
2115         ret = 0;
2116 bail:
2117         ocfs2_free_dir_lookup_result(&lookup);
2118
2119         mlog_exit(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, 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
2292         ocfs2_journal_dirty(handle, di_bh);
2293         if (ret) {
2294                 mlog_errno(ret);
2295                 goto out;
2296         }
2297
2298         i_size_write(inode, size);
2299         inode->i_nlink = 2;
2300         inode->i_blocks = ocfs2_inode_sector_count(inode);
2301
2302         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2303         if (ret < 0)
2304                 mlog_errno(ret);
2305
2306 out:
2307         return ret;
2308 }
2309
2310 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2311                                  handle_t *handle,
2312                                  struct inode *parent,
2313                                  struct inode *inode,
2314                                  struct buffer_head *fe_bh,
2315                                  struct ocfs2_alloc_context *data_ac,
2316                                  struct buffer_head **ret_new_bh)
2317 {
2318         int status;
2319         unsigned int size = osb->sb->s_blocksize;
2320         struct buffer_head *new_bh = NULL;
2321         struct ocfs2_dir_entry *de;
2322
2323         mlog_entry_void();
2324
2325         if (ocfs2_new_dir_wants_trailer(inode))
2326                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2327
2328         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2329                                      data_ac, NULL, &new_bh);
2330         if (status < 0) {
2331                 mlog_errno(status);
2332                 goto bail;
2333         }
2334
2335         ocfs2_set_new_buffer_uptodate(inode, new_bh);
2336
2337         status = ocfs2_journal_access_db(handle, inode, new_bh,
2338                                          OCFS2_JOURNAL_ACCESS_CREATE);
2339         if (status < 0) {
2340                 mlog_errno(status);
2341                 goto bail;
2342         }
2343         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2344
2345         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2346         if (ocfs2_new_dir_wants_trailer(inode)) {
2347                 int size = le16_to_cpu(de->rec_len);
2348
2349                 /*
2350                  * Figure out the size of the hole left over after
2351                  * insertion of '.' and '..'. The trailer wants this
2352                  * information.
2353                  */
2354                 size -= OCFS2_DIR_REC_LEN(2);
2355                 size -= sizeof(struct ocfs2_dir_block_trailer);
2356
2357                 ocfs2_init_dir_trailer(inode, new_bh, size);
2358         }
2359
2360         status = ocfs2_journal_dirty(handle, new_bh);
2361         if (status < 0) {
2362                 mlog_errno(status);
2363                 goto bail;
2364         }
2365
2366         i_size_write(inode, inode->i_sb->s_blocksize);
2367         inode->i_nlink = 2;
2368         inode->i_blocks = ocfs2_inode_sector_count(inode);
2369         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2370         if (status < 0) {
2371                 mlog_errno(status);
2372                 goto bail;
2373         }
2374
2375         status = 0;
2376         if (ret_new_bh) {
2377                 *ret_new_bh = new_bh;
2378                 new_bh = NULL;
2379         }
2380 bail:
2381         brelse(new_bh);
2382
2383         mlog_exit(status);
2384         return status;
2385 }
2386
2387 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2388                                      handle_t *handle, struct inode *dir,
2389                                      struct buffer_head *di_bh,
2390                                      struct buffer_head *dirdata_bh,
2391                                      struct ocfs2_alloc_context *meta_ac,
2392                                      int dx_inline, u32 num_entries,
2393                                      struct buffer_head **ret_dx_root_bh)
2394 {
2395         int ret;
2396         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2397         u16 dr_suballoc_bit;
2398         u64 dr_blkno;
2399         unsigned int num_bits;
2400         struct buffer_head *dx_root_bh = NULL;
2401         struct ocfs2_dx_root_block *dx_root;
2402         struct ocfs2_dir_block_trailer *trailer =
2403                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2404
2405         ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1, &dr_suballoc_bit,
2406                                    &num_bits, &dr_blkno);
2407         if (ret) {
2408                 mlog_errno(ret);
2409                 goto out;
2410         }
2411
2412         mlog(0, "Dir %llu, attach new index block: %llu\n",
2413              (unsigned long long)OCFS2_I(dir)->ip_blkno,
2414              (unsigned long long)dr_blkno);
2415
2416         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2417         if (dx_root_bh == NULL) {
2418                 ret = -EIO;
2419                 goto out;
2420         }
2421         ocfs2_set_new_buffer_uptodate(dir, dx_root_bh);
2422
2423         ret = ocfs2_journal_access_dr(handle, dir, dx_root_bh,
2424                                       OCFS2_JOURNAL_ACCESS_CREATE);
2425         if (ret < 0) {
2426                 mlog_errno(ret);
2427                 goto out;
2428         }
2429
2430         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2431         memset(dx_root, 0, osb->sb->s_blocksize);
2432         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2433         dx_root->dr_suballoc_slot = cpu_to_le16(osb->slot_num);
2434         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2435         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2436         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2437         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2438         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2439         if (le16_to_cpu(trailer->db_free_rec_len))
2440                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2441         else
2442                 dx_root->dr_free_blk = cpu_to_le64(0);
2443
2444         if (dx_inline) {
2445                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2446                 dx_root->dr_entries.de_count =
2447                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2448         } else {
2449                 dx_root->dr_list.l_count =
2450                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2451         }
2452
2453         ret = ocfs2_journal_dirty(handle, dx_root_bh);
2454         if (ret)
2455                 mlog_errno(ret);
2456
2457         ret = ocfs2_journal_access_di(handle, dir, di_bh,
2458                                       OCFS2_JOURNAL_ACCESS_CREATE);
2459         if (ret) {
2460                 mlog_errno(ret);
2461                 goto out;
2462         }
2463
2464         di->i_dx_root = cpu_to_le64(dr_blkno);
2465
2466         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2467         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2468
2469         ret = ocfs2_journal_dirty(handle, di_bh);
2470         if (ret)
2471                 mlog_errno(ret);
2472
2473         *ret_dx_root_bh = dx_root_bh;
2474         dx_root_bh = NULL;
2475
2476 out:
2477         brelse(dx_root_bh);
2478         return ret;
2479 }
2480
2481 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2482                                        handle_t *handle, struct inode *dir,
2483                                        struct buffer_head **dx_leaves,
2484                                        int num_dx_leaves, u64 start_blk)
2485 {
2486         int ret, i;
2487         struct ocfs2_dx_leaf *dx_leaf;
2488         struct buffer_head *bh;
2489
2490         for (i = 0; i < num_dx_leaves; i++) {
2491                 bh = sb_getblk(osb->sb, start_blk + i);
2492                 if (bh == NULL) {
2493                         ret = -EIO;
2494                         goto out;
2495                 }
2496                 dx_leaves[i] = bh;
2497
2498                 ocfs2_set_new_buffer_uptodate(dir, bh);
2499
2500                 ret = ocfs2_journal_access_dl(handle, dir, bh,
2501                                               OCFS2_JOURNAL_ACCESS_CREATE);
2502                 if (ret < 0) {
2503                         mlog_errno(ret);
2504                         goto out;
2505                 }
2506
2507                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2508
2509                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2510                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2511                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2512                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2513                 dx_leaf->dl_list.de_count =
2514                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2515
2516                 mlog(0,
2517                      "Dir %llu, format dx_leaf: %llu, entry count: %u\n",
2518                      (unsigned long long)OCFS2_I(dir)->ip_blkno,
2519                      (unsigned long long)bh->b_blocknr,
2520                      le16_to_cpu(dx_leaf->dl_list.de_count));
2521
2522                 ocfs2_journal_dirty(handle, bh);
2523         }
2524
2525         ret = 0;
2526 out:
2527         return ret;
2528 }
2529
2530 /*
2531  * Allocates and formats a new cluster for use in an indexed dir
2532  * leaf. This version will not do the extent insert, so that it can be
2533  * used by operations which need careful ordering.
2534  */
2535 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2536                                       u32 cpos, handle_t *handle,
2537                                       struct ocfs2_alloc_context *data_ac,
2538                                       struct buffer_head **dx_leaves,
2539                                       int num_dx_leaves, u64 *ret_phys_blkno)
2540 {
2541         int ret;
2542         u32 phys, num;
2543         u64 phys_blkno;
2544         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2545
2546         /*
2547          * XXX: For create, this should claim cluster for the index
2548          * *before* the unindexed insert so that we have a better
2549          * chance of contiguousness as the directory grows in number
2550          * of entries.
2551          */
2552         ret = __ocfs2_claim_clusters(osb, handle, data_ac, 1, 1, &phys, &num);
2553         if (ret) {
2554                 mlog_errno(ret);
2555                 goto out;
2556         }
2557
2558         /*
2559          * Format the new cluster first. That way, we're inserting
2560          * valid data.
2561          */
2562         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2563         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2564                                           num_dx_leaves, phys_blkno);
2565         if (ret) {
2566                 mlog_errno(ret);
2567                 goto out;
2568         }
2569
2570         *ret_phys_blkno = phys_blkno;
2571 out:
2572         return ret;
2573 }
2574
2575 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2576                                     struct ocfs2_extent_tree *et,
2577                                     u32 cpos, handle_t *handle,
2578                                     struct ocfs2_alloc_context *data_ac,
2579                                     struct ocfs2_alloc_context *meta_ac,
2580                                     struct buffer_head **dx_leaves,
2581                                     int num_dx_leaves)
2582 {
2583         int ret;
2584         u64 phys_blkno;
2585         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2586
2587         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2588                                          num_dx_leaves, &phys_blkno);
2589         if (ret) {
2590                 mlog_errno(ret);
2591                 goto out;
2592         }
2593
2594         ret = ocfs2_insert_extent(osb, handle, dir, et, cpos, phys_blkno, 1, 0,
2595                                   meta_ac);
2596         if (ret)
2597                 mlog_errno(ret);
2598 out:
2599         return ret;
2600 }
2601
2602 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2603                                                         int *ret_num_leaves)
2604 {
2605         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2606         struct buffer_head **dx_leaves;
2607
2608         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2609                             GFP_NOFS);
2610         if (dx_leaves && ret_num_leaves)
2611                 *ret_num_leaves = num_dx_leaves;
2612
2613         return dx_leaves;
2614 }
2615
2616 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2617                                  handle_t *handle,
2618                                  struct inode *parent,
2619                                  struct inode *inode,
2620                                  struct buffer_head *di_bh,
2621                                  struct ocfs2_alloc_context *data_ac,
2622                                  struct ocfs2_alloc_context *meta_ac)
2623 {
2624         int ret;
2625         struct buffer_head *leaf_bh = NULL;
2626         struct buffer_head *dx_root_bh = NULL;
2627         struct ocfs2_dx_hinfo hinfo;
2628         struct ocfs2_dx_root_block *dx_root;
2629         struct ocfs2_dx_entry_list *entry_list;
2630
2631         /*
2632          * Our strategy is to create the directory as though it were
2633          * unindexed, then add the index block. This works with very
2634          * little complication since the state of a new directory is a
2635          * very well known quantity.
2636          *
2637          * Essentially, we have two dirents ("." and ".."), in the 1st
2638          * block which need indexing. These are easily inserted into
2639          * the index block.
2640          */
2641
2642         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2643                                     data_ac, &leaf_bh);
2644         if (ret) {
2645                 mlog_errno(ret);
2646                 goto out;
2647         }
2648
2649         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2650                                         meta_ac, 1, 2, &dx_root_bh);
2651         if (ret) {
2652                 mlog_errno(ret);
2653                 goto out;
2654         }
2655         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2656         entry_list = &dx_root->dr_entries;
2657
2658         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2659         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2660         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2661
2662         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2663         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2664
2665 out:
2666         brelse(dx_root_bh);
2667         brelse(leaf_bh);
2668         return ret;
2669 }
2670
2671 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2672                        handle_t *handle,
2673                        struct inode *parent,
2674                        struct inode *inode,
2675                        struct buffer_head *fe_bh,
2676                        struct ocfs2_alloc_context *data_ac,
2677                        struct ocfs2_alloc_context *meta_ac)
2678
2679 {
2680         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2681
2682         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2683                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2684
2685         if (ocfs2_supports_indexed_dirs(osb))
2686                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2687                                              data_ac, meta_ac);
2688
2689         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2690                                      data_ac, NULL);
2691 }
2692
2693 static int ocfs2_dx_dir_index_block(struct inode *dir,
2694                                     handle_t *handle,
2695                                     struct buffer_head **dx_leaves,
2696                                     int num_dx_leaves,
2697                                     u32 *num_dx_entries,
2698                                     struct buffer_head *dirent_bh)
2699 {
2700         int ret = 0, namelen, i;
2701         char *de_buf, *limit;
2702         struct ocfs2_dir_entry *de;
2703         struct buffer_head *dx_leaf_bh;
2704         struct ocfs2_dx_hinfo hinfo;
2705         u64 dirent_blk = dirent_bh->b_blocknr;
2706
2707         de_buf = dirent_bh->b_data;
2708         limit = de_buf + dir->i_sb->s_blocksize;
2709
2710         while (de_buf < limit) {
2711                 de = (struct ocfs2_dir_entry *)de_buf;
2712
2713                 namelen = de->name_len;
2714                 if (!namelen || !de->inode)
2715                         goto inc;
2716
2717                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2718
2719                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2720                 dx_leaf_bh = dx_leaves[i];
2721
2722                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2723                                                  dirent_blk, dx_leaf_bh);
2724                 if (ret) {
2725                         mlog_errno(ret);
2726                         goto out;
2727                 }
2728
2729                 *num_dx_entries = *num_dx_entries + 1;
2730
2731 inc:
2732                 de_buf += le16_to_cpu(de->rec_len);
2733         }
2734
2735 out:
2736         return ret;
2737 }
2738
2739 /*
2740  * XXX: This expects dx_root_bh to already be part of the transaction.
2741  */
2742 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2743                                          struct buffer_head *dx_root_bh,
2744                                          struct buffer_head *dirent_bh)
2745 {
2746         char *de_buf, *limit;
2747         struct ocfs2_dx_root_block *dx_root;
2748         struct ocfs2_dir_entry *de;
2749         struct ocfs2_dx_hinfo hinfo;
2750         u64 dirent_blk = dirent_bh->b_blocknr;
2751
2752         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2753
2754         de_buf = dirent_bh->b_data;
2755         limit = de_buf + dir->i_sb->s_blocksize;
2756
2757         while (de_buf < limit) {
2758                 de = (struct ocfs2_dir_entry *)de_buf;
2759
2760                 if (!de->name_len || !de->inode)
2761                         goto inc;
2762
2763                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2764
2765                 mlog(0,
2766                      "dir: %llu, major: 0x%x minor: 0x%x, index: %u, name: %.*s\n",
2767                      (unsigned long long)dir->i_ino, hinfo.major_hash,
2768                      hinfo.minor_hash,
2769                      le16_to_cpu(dx_root->dr_entries.de_num_used),
2770                      de->name_len, de->name);
2771
2772                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2773                                            dirent_blk);
2774
2775                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2776 inc:
2777                 de_buf += le16_to_cpu(de->rec_len);
2778         }
2779 }
2780
2781 /*
2782  * Count the number of inline directory entries in di_bh and compare
2783  * them against the number of entries we can hold in an inline dx root
2784  * block.
2785  */
2786 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2787                                          struct buffer_head *di_bh)
2788 {
2789         int dirent_count = 0;
2790         char *de_buf, *limit;
2791         struct ocfs2_dir_entry *de;
2792         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2793
2794         de_buf = di->id2.i_data.id_data;
2795         limit = de_buf + i_size_read(dir);
2796
2797         while (de_buf < limit) {
2798                 de = (struct ocfs2_dir_entry *)de_buf;
2799
2800                 if (de->name_len && de->inode)
2801                         dirent_count++;
2802
2803                 de_buf += le16_to_cpu(de->rec_len);
2804         }
2805
2806         /* We are careful to leave room for one extra record. */
2807         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2808 }
2809
2810 /*
2811  * Expand rec_len of the rightmost dirent in a directory block so that it
2812  * contains the end of our valid space for dirents. We do this during
2813  * expansion from an inline directory to one with extents. The first dir block
2814  * in that case is taken from the inline data portion of the inode block.
2815  *
2816  * This will also return the largest amount of contiguous space for a dirent
2817  * in the block. That value is *not* necessarily the last dirent, even after
2818  * expansion. The directory indexing code wants this value for free space
2819  * accounting. We do this here since we're already walking the entire dir
2820  * block.
2821  *
2822  * We add the dir trailer if this filesystem wants it.
2823  */
2824 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2825                                              struct inode *dir)
2826 {
2827         struct super_block *sb = dir->i_sb;
2828         struct ocfs2_dir_entry *de;
2829         struct ocfs2_dir_entry *prev_de;
2830         char *de_buf, *limit;
2831         unsigned int new_size = sb->s_blocksize;
2832         unsigned int bytes, this_hole;
2833         unsigned int largest_hole = 0;
2834
2835         if (ocfs2_new_dir_wants_trailer(dir))
2836                 new_size = ocfs2_dir_trailer_blk_off(sb);
2837
2838         bytes = new_size - old_size;
2839
2840         limit = start + old_size;
2841         de_buf = start;
2842         de = (struct ocfs2_dir_entry *)de_buf;
2843         do {
2844                 this_hole = ocfs2_figure_dirent_hole(de);
2845                 if (this_hole > largest_hole)
2846                         largest_hole = this_hole;
2847
2848                 prev_de = de;
2849                 de_buf += le16_to_cpu(de->rec_len);
2850                 de = (struct ocfs2_dir_entry *)de_buf;
2851         } while (de_buf < limit);
2852
2853         le16_add_cpu(&prev_de->rec_len, bytes);
2854
2855         /* We need to double check this after modification of the final
2856          * dirent. */
2857         this_hole = ocfs2_figure_dirent_hole(prev_de);
2858         if (this_hole > largest_hole)
2859                 largest_hole = this_hole;
2860
2861         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2862                 return largest_hole;
2863         return 0;
2864 }
2865
2866 /*
2867  * We allocate enough clusters to fulfill "blocks_wanted", but set
2868  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2869  * rest automatically for us.
2870  *
2871  * *first_block_bh is a pointer to the 1st data block allocated to the
2872  *  directory.
2873  */
2874 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2875                                    unsigned int blocks_wanted,
2876                                    struct ocfs2_dir_lookup_result *lookup,
2877                                    struct buffer_head **first_block_bh)
2878 {
2879         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2880         struct super_block *sb = dir->i_sb;
2881         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2882                 credits = ocfs2_inline_to_extents_credits(sb);
2883         u64 dx_insert_blkno, blkno,
2884                 bytes = blocks_wanted << sb->s_blocksize_bits;
2885         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2886         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2887         struct ocfs2_alloc_context *data_ac;
2888         struct ocfs2_alloc_context *meta_ac = NULL;
2889         struct buffer_head *dirdata_bh = NULL;
2890         struct buffer_head *dx_root_bh = NULL;
2891         struct buffer_head **dx_leaves = NULL;
2892         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2893         handle_t *handle;
2894         struct ocfs2_extent_tree et;
2895         struct ocfs2_extent_tree dx_et;
2896         int did_quota = 0, bytes_allocated = 0;
2897
2898         ocfs2_init_dinode_extent_tree(&et, dir, di_bh);
2899
2900         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2901         dx_alloc = 0;
2902
2903         down_write(&oi->ip_alloc_sem);
2904
2905         if (ocfs2_supports_indexed_dirs(osb)) {
2906                 credits += ocfs2_add_dir_index_credits(sb);
2907
2908                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2909                 if (!dx_inline) {
2910                         /* Add one more cluster for an index leaf */
2911                         dx_alloc++;
2912                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2913                                                                 &num_dx_leaves);
2914                         if (!dx_leaves) {
2915                                 ret = -ENOMEM;
2916                                 mlog_errno(ret);
2917                                 goto out;
2918                         }
2919                 }
2920
2921                 /* This gets us the dx_root */
2922                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2923                 if (ret) {
2924                         mlog_errno(ret);
2925                         goto out;
2926                 }
2927         }
2928
2929         /*
2930          * We should never need more than 2 clusters for the unindexed
2931          * tree - maximum dirent size is far less than one block. In
2932          * fact, the only time we'd need more than one cluster is if
2933          * blocksize == clustersize and the dirent won't fit in the
2934          * extra space that the expansion to a single block gives. As
2935          * of today, that only happens on 4k/4k file systems.
2936          */
2937         BUG_ON(alloc > 2);
2938
2939         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2940         if (ret) {
2941                 mlog_errno(ret);
2942                 goto out;
2943         }
2944
2945         /*
2946          * Prepare for worst case allocation scenario of two separate
2947          * extents in the unindexed tree.
2948          */
2949         if (alloc == 2)
2950                 credits += OCFS2_SUBALLOC_ALLOC;
2951
2952         handle = ocfs2_start_trans(osb, credits);
2953         if (IS_ERR(handle)) {
2954                 ret = PTR_ERR(handle);
2955                 mlog_errno(ret);
2956                 goto out;
2957         }
2958
2959         if (vfs_dq_alloc_space_nodirty(dir,
2960                                 ocfs2_clusters_to_bytes(osb->sb,
2961                                                         alloc + dx_alloc))) {
2962                 ret = -EDQUOT;
2963                 goto out_commit;
2964         }
2965         did_quota = 1;
2966
2967         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2968                 /*
2969                  * Allocate our index cluster first, to maximize the
2970                  * possibility that unindexed leaves grow
2971                  * contiguously.
2972                  */
2973                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2974                                                  dx_leaves, num_dx_leaves,
2975                                                  &dx_insert_blkno);
2976                 if (ret) {
2977                         mlog_errno(ret);
2978                         goto out_commit;
2979                 }
2980                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2981         }
2982
2983         /*
2984          * Try to claim as many clusters as the bitmap can give though
2985          * if we only get one now, that's enough to continue. The rest
2986          * will be claimed after the conversion to extents.
2987          */
2988         ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off, &len);
2989         if (ret) {
2990                 mlog_errno(ret);
2991                 goto out_commit;
2992         }
2993         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2994
2995         /*
2996          * Operations are carefully ordered so that we set up the new
2997          * data block first. The conversion from inline data to
2998          * extents follows.
2999          */
3000         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3001         dirdata_bh = sb_getblk(sb, blkno);
3002         if (!dirdata_bh) {
3003                 ret = -EIO;
3004                 mlog_errno(ret);
3005                 goto out_commit;
3006         }
3007
3008         ocfs2_set_new_buffer_uptodate(dir, dirdata_bh);
3009
3010         ret = ocfs2_journal_access_db(handle, dir, dirdata_bh,
3011                                       OCFS2_JOURNAL_ACCESS_CREATE);
3012         if (ret) {
3013                 mlog_errno(ret);
3014                 goto out_commit;
3015         }
3016
3017         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
3018         memset(dirdata_bh->b_data + i_size_read(dir), 0,
3019                sb->s_blocksize - i_size_read(dir));
3020         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
3021         if (ocfs2_new_dir_wants_trailer(dir)) {
3022                 /*
3023                  * Prepare the dir trailer up front. It will otherwise look
3024                  * like a valid dirent. Even if inserting the index fails
3025                  * (unlikely), then all we'll have done is given first dir
3026                  * block a small amount of fragmentation.
3027                  */
3028                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
3029         }
3030
3031         ret = ocfs2_journal_dirty(handle, dirdata_bh);
3032         if (ret) {
3033                 mlog_errno(ret);
3034                 goto out_commit;
3035         }
3036
3037         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
3038                 /*
3039                  * Dx dirs with an external cluster need to do this up
3040                  * front. Inline dx root's get handled later, after
3041                  * we've allocated our root block. We get passed back
3042                  * a total number of items so that dr_num_entries can
3043                  * be correctly set once the dx_root has been
3044                  * allocated.
3045                  */
3046                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
3047                                                num_dx_leaves, &num_dx_entries,
3048                                                dirdata_bh);
3049                 if (ret) {
3050                         mlog_errno(ret);
3051                         goto out_commit;
3052                 }
3053         }
3054
3055         /*
3056          * Set extent, i_size, etc on the directory. After this, the
3057          * inode should contain the same exact dirents as before and
3058          * be fully accessible from system calls.
3059          *
3060          * We let the later dirent insert modify c/mtime - to the user
3061          * the data hasn't changed.
3062          */
3063         ret = ocfs2_journal_access_di(handle, dir, di_bh,
3064                                       OCFS2_JOURNAL_ACCESS_CREATE);
3065         if (ret) {
3066                 mlog_errno(ret);
3067                 goto out_commit;
3068         }
3069
3070         spin_lock(&oi->ip_lock);
3071         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
3072         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
3073         spin_unlock(&oi->ip_lock);
3074
3075         ocfs2_dinode_new_extent_list(dir, di);
3076
3077         i_size_write(dir, sb->s_blocksize);
3078         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3079
3080         di->i_size = cpu_to_le64(sb->s_blocksize);
3081         di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3082         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3083
3084         /*
3085          * This should never fail as our extent list is empty and all
3086          * related blocks have been journaled already.
3087          */
3088         ret = ocfs2_insert_extent(osb, handle, dir, &et, 0, blkno, len,
3089                                   0, NULL);
3090         if (ret) {
3091                 mlog_errno(ret);
3092                 goto out_commit;
3093         }
3094
3095         /*
3096          * Set i_blocks after the extent insert for the most up to
3097          * date ip_clusters value.
3098          */
3099         dir->i_blocks = ocfs2_inode_sector_count(dir);
3100
3101         ret = ocfs2_journal_dirty(handle, di_bh);
3102         if (ret) {
3103                 mlog_errno(ret);
3104                 goto out_commit;
3105         }
3106
3107         if (ocfs2_supports_indexed_dirs(osb)) {
3108                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3109                                                 dirdata_bh, meta_ac, dx_inline,
3110                                                 num_dx_entries, &dx_root_bh);
3111                 if (ret) {
3112                         mlog_errno(ret);
3113                         goto out_commit;
3114                 }
3115
3116                 if (dx_inline) {
3117                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3118                                                       dirdata_bh);
3119                 } else {
3120                         ocfs2_init_dx_root_extent_tree(&dx_et, dir, dx_root_bh);
3121                         ret = ocfs2_insert_extent(osb, handle, dir, &dx_et, 0,
3122                                                   dx_insert_blkno, 1, 0, NULL);
3123                         if (ret)
3124                                 mlog_errno(ret);
3125                 }
3126         }
3127
3128         /*
3129          * We asked for two clusters, but only got one in the 1st
3130          * pass. Claim the 2nd cluster as a separate extent.
3131          */
3132         if (alloc > len) {
3133                 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
3134                                            &len);
3135                 if (ret) {
3136                         mlog_errno(ret);
3137                         goto out_commit;
3138                 }
3139                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3140
3141                 ret = ocfs2_insert_extent(osb, handle, dir, &et, 1,
3142                                           blkno, len, 0, NULL);
3143                 if (ret) {
3144                         mlog_errno(ret);
3145                         goto out_commit;
3146                 }
3147                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3148         }
3149
3150         *first_block_bh = dirdata_bh;
3151         dirdata_bh = NULL;
3152         if (ocfs2_supports_indexed_dirs(osb)) {
3153                 unsigned int off;
3154
3155                 if (!dx_inline) {
3156                         /*
3157                          * We need to return the correct block within the
3158                          * cluster which should hold our entry.
3159                          */
3160                         off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3161                                                     &lookup->dl_hinfo);
3162                         get_bh(dx_leaves[off]);
3163                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3164                 }
3165                 lookup->dl_dx_root_bh = dx_root_bh;
3166                 dx_root_bh = NULL;
3167         }
3168
3169 out_commit:
3170         if (ret < 0 && did_quota)
3171                 vfs_dq_free_space_nodirty(dir, bytes_allocated);
3172
3173         ocfs2_commit_trans(osb, handle);
3174
3175 out:
3176         up_write(&oi->ip_alloc_sem);
3177         if (data_ac)
3178                 ocfs2_free_alloc_context(data_ac);
3179         if (meta_ac)
3180                 ocfs2_free_alloc_context(meta_ac);
3181
3182         if (dx_leaves) {
3183                 for (i = 0; i < num_dx_leaves; i++)
3184                         brelse(dx_leaves[i]);
3185                 kfree(dx_leaves);
3186         }
3187
3188         brelse(dirdata_bh);
3189         brelse(dx_root_bh);
3190
3191         return ret;
3192 }
3193
3194 /* returns a bh of the 1st new block in the allocation. */
3195 static int ocfs2_do_extend_dir(struct super_block *sb,
3196                                handle_t *handle,
3197                                struct inode *dir,
3198                                struct buffer_head *parent_fe_bh,
3199                                struct ocfs2_alloc_context *data_ac,
3200                                struct ocfs2_alloc_context *meta_ac,
3201                                struct buffer_head **new_bh)
3202 {
3203         int status;
3204         int extend, did_quota = 0;
3205         u64 p_blkno, v_blkno;
3206
3207         spin_lock(&OCFS2_I(dir)->ip_lock);
3208         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3209         spin_unlock(&OCFS2_I(dir)->ip_lock);
3210
3211         if (extend) {
3212                 u32 offset = OCFS2_I(dir)->ip_clusters;
3213
3214                 if (vfs_dq_alloc_space_nodirty(dir,
3215                                         ocfs2_clusters_to_bytes(sb, 1))) {
3216                         status = -EDQUOT;
3217                         goto bail;
3218                 }
3219                 did_quota = 1;
3220
3221                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3222                                               1, 0, parent_fe_bh, handle,
3223                                               data_ac, meta_ac, NULL);
3224                 BUG_ON(status == -EAGAIN);
3225                 if (status < 0) {
3226                         mlog_errno(status);
3227                         goto bail;
3228                 }
3229         }
3230
3231         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3232         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3233         if (status < 0) {
3234                 mlog_errno(status);
3235                 goto bail;
3236         }
3237
3238         *new_bh = sb_getblk(sb, p_blkno);
3239         if (!*new_bh) {
3240                 status = -EIO;
3241                 mlog_errno(status);
3242                 goto bail;
3243         }
3244         status = 0;
3245 bail:
3246         if (did_quota && status < 0)
3247                 vfs_dq_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3248         mlog_exit(status);
3249         return status;
3250 }
3251
3252 /*
3253  * Assumes you already have a cluster lock on the directory.
3254  *
3255  * 'blocks_wanted' is only used if we have an inline directory which
3256  * is to be turned into an extent based one. The size of the dirent to
3257  * insert might be larger than the space gained by growing to just one
3258  * block, so we may have to grow the inode by two blocks in that case.
3259  *
3260  * If the directory is already indexed, dx_root_bh must be provided.
3261  */
3262 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3263                             struct inode *dir,
3264                             struct buffer_head *parent_fe_bh,
3265                             unsigned int blocks_wanted,
3266                             struct ocfs2_dir_lookup_result *lookup,
3267                             struct buffer_head **new_de_bh)
3268 {
3269         int status = 0;
3270         int credits, num_free_extents, drop_alloc_sem = 0;
3271         loff_t dir_i_size;
3272         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3273         struct ocfs2_extent_list *el = &fe->id2.i_list;
3274         struct ocfs2_alloc_context *data_ac = NULL;
3275         struct ocfs2_alloc_context *meta_ac = NULL;
3276         handle_t *handle = NULL;
3277         struct buffer_head *new_bh = NULL;
3278         struct ocfs2_dir_entry * de;
3279         struct super_block *sb = osb->sb;
3280         struct ocfs2_extent_tree et;
3281         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3282
3283         mlog_entry_void();
3284
3285         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3286                 /*
3287                  * This would be a code error as an inline directory should
3288                  * never have an index root.
3289                  */
3290                 BUG_ON(dx_root_bh);
3291
3292                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3293                                                  blocks_wanted, lookup,
3294                                                  &new_bh);
3295                 if (status) {
3296                         mlog_errno(status);
3297                         goto bail;
3298                 }
3299
3300                 /* Expansion from inline to an indexed directory will
3301                  * have given us this. */
3302                 dx_root_bh = lookup->dl_dx_root_bh;
3303
3304                 if (blocks_wanted == 1) {
3305                         /*
3306                          * If the new dirent will fit inside the space
3307                          * created by pushing out to one block, then
3308                          * we can complete the operation
3309                          * here. Otherwise we have to expand i_size
3310                          * and format the 2nd block below.
3311                          */
3312                         BUG_ON(new_bh == NULL);
3313                         goto bail_bh;
3314                 }
3315
3316                 /*
3317                  * Get rid of 'new_bh' - we want to format the 2nd
3318                  * data block and return that instead.
3319                  */
3320                 brelse(new_bh);
3321                 new_bh = NULL;
3322
3323                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3324                 drop_alloc_sem = 1;
3325                 dir_i_size = i_size_read(dir);
3326                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3327                 goto do_extend;
3328         }
3329
3330         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3331         drop_alloc_sem = 1;
3332         dir_i_size = i_size_read(dir);
3333         mlog(0, "extending dir %llu (i_size = %lld)\n",
3334              (unsigned long long)OCFS2_I(dir)->ip_blkno, dir_i_size);
3335
3336         /* dir->i_size is always block aligned. */
3337         spin_lock(&OCFS2_I(dir)->ip_lock);
3338         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3339                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3340                 ocfs2_init_dinode_extent_tree(&et, dir, parent_fe_bh);
3341                 num_free_extents = ocfs2_num_free_extents(osb, dir, &et);
3342                 if (num_free_extents < 0) {
3343                         status = num_free_extents;
3344                         mlog_errno(status);
3345                         goto bail;
3346                 }
3347
3348                 if (!num_free_extents) {
3349                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3350                         if (status < 0) {
3351                                 if (status != -ENOSPC)
3352                                         mlog_errno(status);
3353                                 goto bail;
3354                         }
3355                 }
3356
3357                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3358                 if (status < 0) {
3359                         if (status != -ENOSPC)
3360                                 mlog_errno(status);
3361                         goto bail;
3362                 }
3363
3364                 credits = ocfs2_calc_extend_credits(sb, el, 1);
3365         } else {
3366                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3367                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3368         }
3369
3370 do_extend:
3371         if (ocfs2_dir_indexed(dir))
3372                 credits++; /* For attaching the new dirent block to the
3373                             * dx_root */
3374
3375         handle = ocfs2_start_trans(osb, credits);
3376         if (IS_ERR(handle)) {
3377                 status = PTR_ERR(handle);
3378                 handle = NULL;
3379                 mlog_errno(status);
3380                 goto bail;
3381         }
3382
3383         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3384                                      data_ac, meta_ac, &new_bh);
3385         if (status < 0) {
3386                 mlog_errno(status);
3387                 goto bail;
3388         }
3389
3390         ocfs2_set_new_buffer_uptodate(dir, new_bh);
3391
3392         status = ocfs2_journal_access_db(handle, dir, new_bh,
3393                                          OCFS2_JOURNAL_ACCESS_CREATE);
3394         if (status < 0) {
3395                 mlog_errno(status);
3396                 goto bail;
3397         }
3398         memset(new_bh->b_data, 0, sb->s_blocksize);
3399
3400         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3401         de->inode = 0;
3402         if (ocfs2_supports_dir_trailer(dir)) {
3403                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3404
3405                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3406
3407                 if (ocfs2_dir_indexed(dir)) {
3408                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3409                                                            dx_root_bh, new_bh);
3410                         if (status) {
3411                                 mlog_errno(status);
3412                                 goto bail;
3413                         }
3414                 }
3415         } else {
3416                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3417         }
3418         status = ocfs2_journal_dirty(handle, new_bh);
3419         if (status < 0) {
3420                 mlog_errno(status);
3421                 goto bail;
3422         }
3423
3424         dir_i_size += dir->i_sb->s_blocksize;
3425         i_size_write(dir, dir_i_size);
3426         dir->i_blocks = ocfs2_inode_sector_count(dir);
3427         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3428         if (status < 0) {
3429                 mlog_errno(status);
3430                 goto bail;
3431         }
3432
3433 bail_bh:
3434         *new_de_bh = new_bh;
3435         get_bh(*new_de_bh);
3436 bail:
3437         if (handle)
3438                 ocfs2_commit_trans(osb, handle);
3439         if (drop_alloc_sem)
3440                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3441
3442         if (data_ac)
3443                 ocfs2_free_alloc_context(data_ac);
3444         if (meta_ac)
3445                 ocfs2_free_alloc_context(meta_ac);
3446
3447         brelse(new_bh);
3448
3449         mlog_exit(status);
3450         return status;
3451 }
3452
3453 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3454                                    const char *name, int namelen,
3455                                    struct buffer_head **ret_de_bh,
3456                                    unsigned int *blocks_wanted)
3457 {
3458         int ret;
3459         struct super_block *sb = dir->i_sb;
3460         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3461         struct ocfs2_dir_entry *de, *last_de = NULL;
3462         char *de_buf, *limit;
3463         unsigned long offset = 0;
3464         unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3465
3466         /*
3467          * This calculates how many free bytes we'd have in block zero, should
3468          * this function force expansion to an extent tree.
3469          */
3470         if (ocfs2_new_dir_wants_trailer(dir))
3471                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3472         else
3473                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3474
3475         de_buf = di->id2.i_data.id_data;
3476         limit = de_buf + i_size_read(dir);
3477         rec_len = OCFS2_DIR_REC_LEN(namelen);
3478
3479         while (de_buf < limit) {
3480                 de = (struct ocfs2_dir_entry *)de_buf;
3481
3482                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3483                         ret = -ENOENT;
3484                         goto out;
3485                 }
3486                 if (ocfs2_match(namelen, name, de)) {
3487                         ret = -EEXIST;
3488                         goto out;
3489                 }
3490                 /*
3491                  * No need to check for a trailing dirent record here as
3492                  * they're not used for inline dirs.
3493                  */
3494
3495                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3496                         /* Ok, we found a spot. Return this bh and let
3497                          * the caller actually fill it in. */
3498                         *ret_de_bh = di_bh;
3499                         get_bh(*ret_de_bh);
3500                         ret = 0;
3501                         goto out;
3502                 }
3503
3504                 last_de = de;
3505                 de_buf += le16_to_cpu(de->rec_len);
3506                 offset += le16_to_cpu(de->rec_len);
3507         }
3508
3509         /*
3510          * We're going to require expansion of the directory - figure
3511          * out how many blocks we'll need so that a place for the
3512          * dirent can be found.
3513          */
3514         *blocks_wanted = 1;
3515         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3516         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3517                 *blocks_wanted = 2;
3518
3519         ret = -ENOSPC;
3520 out:
3521         return ret;
3522 }
3523
3524 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3525                                    int namelen, struct buffer_head **ret_de_bh)
3526 {
3527         unsigned long offset;
3528         struct buffer_head *bh = NULL;
3529         unsigned short rec_len;
3530         struct ocfs2_dir_entry *de;
3531         struct super_block *sb = dir->i_sb;
3532         int status;
3533         int blocksize = dir->i_sb->s_blocksize;
3534
3535         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3536         if (status) {
3537                 mlog_errno(status);
3538                 goto bail;
3539         }
3540
3541         rec_len = OCFS2_DIR_REC_LEN(namelen);
3542         offset = 0;
3543         de = (struct ocfs2_dir_entry *) bh->b_data;
3544         while (1) {
3545                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3546                         brelse(bh);
3547                         bh = NULL;
3548
3549                         if (i_size_read(dir) <= offset) {
3550                                 /*
3551                                  * Caller will have to expand this
3552                                  * directory.
3553                                  */
3554                                 status = -ENOSPC;
3555                                 goto bail;
3556                         }
3557                         status = ocfs2_read_dir_block(dir,
3558                                              offset >> sb->s_blocksize_bits,
3559                                              &bh, 0);
3560                         if (status) {
3561                                 mlog_errno(status);
3562                                 goto bail;
3563                         }
3564                         /* move to next block */
3565                         de = (struct ocfs2_dir_entry *) bh->b_data;
3566                 }
3567                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3568                         status = -ENOENT;
3569                         goto bail;
3570                 }
3571                 if (ocfs2_match(namelen, name, de)) {
3572                         status = -EEXIST;
3573                         goto bail;
3574                 }
3575
3576                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3577                                            blocksize))
3578                         goto next;
3579
3580                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3581                         /* Ok, we found a spot. Return this bh and let
3582                          * the caller actually fill it in. */
3583                         *ret_de_bh = bh;
3584                         get_bh(*ret_de_bh);
3585                         status = 0;
3586                         goto bail;
3587                 }
3588 next:
3589                 offset += le16_to_cpu(de->rec_len);
3590                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3591         }
3592
3593         status = 0;
3594 bail:
3595         brelse(bh);
3596
3597         mlog_exit(status);
3598         return status;
3599 }
3600
3601 static int dx_leaf_sort_cmp(const void *a, const void *b)
3602 {
3603         const struct ocfs2_dx_entry *entry1 = a;
3604         const struct ocfs2_dx_entry *entry2 = b;
3605         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3606         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3607         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3608         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3609
3610         if (major_hash1 > major_hash2)
3611                 return 1;
3612         if (major_hash1 < major_hash2)
3613                 return -1;
3614
3615         /*
3616          * It is not strictly necessary to sort by minor
3617          */
3618         if (minor_hash1 > minor_hash2)
3619                 return 1;
3620         if (minor_hash1 < minor_hash2)
3621                 return -1;
3622         return 0;
3623 }
3624
3625 static void dx_leaf_sort_swap(void *a, void *b, int size)
3626 {
3627         struct ocfs2_dx_entry *entry1 = a;
3628         struct ocfs2_dx_entry *entry2 = b;
3629         struct ocfs2_dx_entry tmp;
3630
3631         BUG_ON(size != sizeof(*entry1));
3632
3633         tmp = *entry1;
3634         *entry1 = *entry2;
3635         *entry2 = tmp;
3636 }
3637
3638 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3639 {
3640         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3641         int i, num = le16_to_cpu(dl_list->de_num_used);
3642
3643         for (i = 0; i < (num - 1); i++) {
3644                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3645                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3646                         return 0;
3647         }
3648
3649         return 1;
3650 }
3651
3652 /*
3653  * Find the optimal value to split this leaf on. This expects the leaf
3654  * entries to be in sorted order.
3655  *
3656  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3657  * the hash we want to insert.
3658  *
3659  * This function is only concerned with the major hash - that which
3660  * determines which cluster an item belongs to.
3661  */
3662 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3663                                         u32 leaf_cpos, u32 insert_hash,
3664                                         u32 *split_hash)
3665 {
3666         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3667         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3668         int allsame;
3669
3670         /*
3671          * There's a couple rare, but nasty corner cases we have to
3672          * check for here. All of them involve a leaf where all value
3673          * have the same hash, which is what we look for first.
3674          *
3675          * Most of the time, all of the above is false, and we simply
3676          * pick the median value for a split.
3677          */
3678         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3679         if (allsame) {
3680                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3681
3682                 if (val == insert_hash) {
3683                         /*
3684                          * No matter where we would choose to split,
3685                          * the new entry would want to occupy the same
3686                          * block as these. Since there's no space left
3687                          * in their existing block, we know there
3688                          * won't be space after the split.
3689                          */
3690                         return -ENOSPC;
3691                 }
3692
3693                 if (val == leaf_cpos) {
3694                         /*
3695                          * Because val is the same as leaf_cpos (which
3696                          * is the smallest value this leaf can have),
3697                          * yet is not equal to insert_hash, then we
3698                          * know that insert_hash *must* be larger than
3699                          * val (and leaf_cpos). At least cpos+1 in value.
3700                          *
3701                          * We also know then, that there cannot be an
3702                          * adjacent extent (otherwise we'd be looking
3703                          * at it). Choosing this value gives us a
3704                          * chance to get some contiguousness.
3705                          */
3706                         *split_hash = leaf_cpos + 1;
3707                         return 0;
3708                 }
3709
3710                 if (val > insert_hash) {
3711                         /*
3712                          * val can not be the same as insert hash, and
3713                          * also must be larger than leaf_cpos. Also,
3714                          * we know that there can't be a leaf between
3715                          * cpos and val, otherwise the entries with
3716                          * hash 'val' would be there.
3717                          */
3718                         *split_hash = val;
3719                         return 0;
3720                 }
3721
3722                 *split_hash = insert_hash;
3723                 return 0;
3724         }
3725
3726         /*
3727          * Since the records are sorted and the checks above
3728          * guaranteed that not all records in this block are the same,
3729          * we simple travel forward, from the median, and pick the 1st
3730          * record whose value is larger than leaf_cpos.
3731          */
3732         for (i = (num_used / 2); i < num_used; i++)
3733                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3734                     leaf_cpos)
3735                         break;
3736
3737         BUG_ON(i == num_used); /* Should be impossible */
3738         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3739         return 0;
3740 }
3741
3742 /*
3743  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3744  * larger than split_hash into new_dx_leaves. We use a temporary
3745  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3746  *
3747  * Since the block offset inside a leaf (cluster) is a constant mask
3748  * of minor_hash, we can optimize - an item at block offset X within
3749  * the original cluster, will be at offset X within the new cluster.
3750  */
3751 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3752                                        handle_t *handle,
3753                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3754                                        struct buffer_head **orig_dx_leaves,
3755                                        struct buffer_head **new_dx_leaves,
3756                                        int num_dx_leaves)
3757 {
3758         int i, j, num_used;
3759         u32 major_hash;
3760         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3761         struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3762         struct ocfs2_dx_entry *dx_entry;
3763
3764         tmp_list = &tmp_dx_leaf->dl_list;
3765
3766         for (i = 0; i < num_dx_leaves; i++) {
3767                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3768                 orig_list = &orig_dx_leaf->dl_list;
3769                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3770                 new_list = &new_dx_leaf->dl_list;
3771
3772                 num_used = le16_to_cpu(orig_list->de_num_used);
3773
3774                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3775                 tmp_list->de_num_used = cpu_to_le16(0);
3776                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3777
3778                 for (j = 0; j < num_used; j++) {
3779                         dx_entry = &orig_list->de_entries[j];
3780                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3781                         if (major_hash >= split_hash)
3782                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3783                                                               dx_entry);
3784                         else
3785                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3786                                                               dx_entry);
3787                 }
3788                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3789
3790                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3791                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3792         }
3793 }
3794
3795 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3796                                           struct ocfs2_dx_root_block *dx_root)
3797 {
3798         int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3799
3800         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list, 1);
3801         credits += ocfs2_quota_trans_credits(osb->sb);
3802         return credits;
3803 }
3804
3805 /*
3806  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3807  * half our entries into.
3808  */
3809 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3810                                   struct buffer_head *dx_root_bh,
3811                                   struct buffer_head *dx_leaf_bh,
3812                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3813                                   u64 leaf_blkno)
3814 {
3815         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3816         int credits, ret, i, num_used, did_quota = 0;
3817         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3818         u64 orig_leaves_start;
3819         int num_dx_leaves;
3820         struct buffer_head **orig_dx_leaves = NULL;
3821         struct buffer_head **new_dx_leaves = NULL;
3822         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3823         struct ocfs2_extent_tree et;
3824         handle_t *handle = NULL;
3825         struct ocfs2_dx_root_block *dx_root;
3826         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3827
3828         mlog(0, "DX Dir: %llu, rebalance leaf leaf_blkno: %llu insert: %u\n",
3829              (unsigned long long)OCFS2_I(dir)->ip_blkno,
3830              (unsigned long long)leaf_blkno, insert_hash);
3831
3832         ocfs2_init_dx_root_extent_tree(&et, dir, dx_root_bh);
3833
3834         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3835         /*
3836          * XXX: This is a rather large limit. We should use a more
3837          * realistic value.
3838          */
3839         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3840                 return -ENOSPC;
3841
3842         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3843         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3844                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3845                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3846                      (unsigned long long)leaf_blkno, num_used);
3847                 ret = -EIO;
3848                 goto out;
3849         }
3850
3851         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3852         if (!orig_dx_leaves) {
3853                 ret = -ENOMEM;
3854                 mlog_errno(ret);
3855                 goto out;
3856         }
3857
3858         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3859         if (!new_dx_leaves) {
3860                 ret = -ENOMEM;
3861                 mlog_errno(ret);
3862                 goto out;
3863         }
3864
3865         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3866         if (ret) {
3867                 if (ret != -ENOSPC)
3868                         mlog_errno(ret);
3869                 goto out;
3870         }
3871
3872         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3873         handle = ocfs2_start_trans(osb, credits);
3874         if (IS_ERR(handle)) {
3875                 ret = PTR_ERR(handle);
3876                 handle = NULL;
3877                 mlog_errno(ret);
3878                 goto out;
3879         }
3880
3881         if (vfs_dq_alloc_space_nodirty(dir,
3882                                        ocfs2_clusters_to_bytes(dir->i_sb, 1))) {
3883                 ret = -EDQUOT;
3884                 goto out_commit;
3885         }
3886         did_quota = 1;
3887
3888         ret = ocfs2_journal_access_dl(handle, dir, dx_leaf_bh,
3889                                       OCFS2_JOURNAL_ACCESS_WRITE);
3890         if (ret) {
3891                 mlog_errno(ret);
3892                 goto out_commit;
3893         }
3894
3895         /*
3896          * This block is changing anyway, so we can sort it in place.
3897          */
3898         sort(dx_leaf->dl_list.de_entries, num_used,
3899              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3900              dx_leaf_sort_swap);
3901
3902         ret = ocfs2_journal_dirty(handle, dx_leaf_bh);
3903         if (ret) {
3904                 mlog_errno(ret);
3905                 goto out_commit;
3906         }
3907
3908         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3909                                            &split_hash);
3910         if (ret) {
3911                 mlog_errno(ret);
3912                 goto  out_commit;
3913         }
3914
3915         mlog(0, "Split leaf (%u) at %u, insert major hash is %u\n",
3916              leaf_cpos, split_hash, insert_hash);
3917
3918         /*
3919          * We have to carefully order operations here. There are items
3920          * which want to be in the new cluster before insert, but in
3921          * order to put those items in the new cluster, we alter the
3922          * old cluster. A failure to insert gets nasty.
3923          *
3924          * So, start by reserving writes to the old
3925          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3926          * the new cluster for us, before inserting it. The insert
3927          * won't happen if there's an error before that. Once the
3928          * insert is done then, we can transfer from one leaf into the
3929          * other without fear of hitting any error.
3930          */
3931
3932         /*
3933          * The leaf transfer wants some scratch space so that we don't
3934          * wind up doing a bunch of expensive memmove().
3935          */
3936         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3937         if (!tmp_dx_leaf) {
3938                 ret = -ENOMEM;
3939                 mlog_errno(ret);
3940                 goto out_commit;
3941         }
3942
3943         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3944         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3945                                    orig_dx_leaves);
3946         if (ret) {
3947                 mlog_errno(ret);
3948                 goto out_commit;
3949         }
3950
3951         for (i = 0; i < num_dx_leaves; i++) {
3952                 ret = ocfs2_journal_access_dl(handle, dir, orig_dx_leaves[i],
3953                                               OCFS2_JOURNAL_ACCESS_WRITE);
3954                 if (ret) {
3955                         mlog_errno(ret);
3956                         goto out_commit;
3957                 }
3958         }
3959
3960         cpos = split_hash;
3961         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3962                                        data_ac, meta_ac, new_dx_leaves,
3963                                        num_dx_leaves);
3964         if (ret) {
3965                 mlog_errno(ret);
3966                 goto out_commit;
3967         }
3968
3969         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3970                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3971
3972 out_commit:
3973         if (ret < 0 && did_quota)
3974                 vfs_dq_free_space_nodirty(dir,
3975                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3976
3977         ocfs2_commit_trans(osb, handle);
3978
3979 out:
3980         if (orig_dx_leaves || new_dx_leaves) {
3981                 for (i = 0; i < num_dx_leaves; i++) {
3982                         if (orig_dx_leaves)
3983                                 brelse(orig_dx_leaves[i]);
3984                         if (new_dx_leaves)
3985                                 brelse(new_dx_leaves[i]);
3986                 }
3987                 kfree(orig_dx_leaves);
3988                 kfree(new_dx_leaves);
3989         }
3990
3991         if (meta_ac)
3992                 ocfs2_free_alloc_context(meta_ac);
3993         if (data_ac)
3994                 ocfs2_free_alloc_context(data_ac);
3995
3996         kfree(tmp_dx_leaf);
3997         return ret;
3998 }
3999
4000 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
4001                                    struct buffer_head *di_bh,
4002                                    struct buffer_head *dx_root_bh,
4003                                    const char *name, int namelen,
4004                                    struct ocfs2_dir_lookup_result *lookup)
4005 {
4006         int ret, rebalanced = 0;
4007         struct ocfs2_dx_root_block *dx_root;
4008         struct buffer_head *dx_leaf_bh = NULL;
4009         struct ocfs2_dx_leaf *dx_leaf;
4010         u64 blkno;
4011         u32 leaf_cpos;
4012
4013         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4014
4015 restart_search:
4016         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
4017                                   &leaf_cpos, &blkno);
4018         if (ret) {
4019                 mlog_errno(ret);
4020                 goto out;
4021         }
4022
4023         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
4024         if (ret) {
4025                 mlog_errno(ret);
4026                 goto out;
4027         }
4028
4029         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
4030
4031         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
4032             le16_to_cpu(dx_leaf->dl_list.de_count)) {
4033                 if (rebalanced) {
4034                         /*
4035                          * Rebalancing should have provided us with
4036                          * space in an appropriate leaf.
4037                          *
4038                          * XXX: Is this an abnormal condition then?
4039                          * Should we print a message here?
4040                          */
4041                         ret = -ENOSPC;
4042                         goto out;
4043                 }
4044
4045                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
4046                                              &lookup->dl_hinfo, leaf_cpos,
4047                                              blkno);
4048                 if (ret) {
4049                         if (ret != -ENOSPC)
4050                                 mlog_errno(ret);
4051                         goto out;
4052                 }
4053
4054                 /*
4055                  * Restart the lookup. The rebalance might have
4056                  * changed which block our item fits into. Mark our
4057                  * progress, so we only execute this once.
4058                  */
4059                 brelse(dx_leaf_bh);
4060                 dx_leaf_bh = NULL;
4061                 rebalanced = 1;
4062                 goto restart_search;
4063         }
4064
4065         lookup->dl_dx_leaf_bh = dx_leaf_bh;
4066         dx_leaf_bh = NULL;
4067
4068 out:
4069         brelse(dx_leaf_bh);
4070         return ret;
4071 }
4072
4073 static int ocfs2_search_dx_free_list(struct inode *dir,
4074                                      struct buffer_head *dx_root_bh,
4075                                      int namelen,
4076                                      struct ocfs2_dir_lookup_result *lookup)
4077 {
4078         int ret = -ENOSPC;
4079         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4080         struct ocfs2_dir_block_trailer *db;
4081         u64 next_block;
4082         int rec_len = OCFS2_DIR_REC_LEN(namelen);
4083         struct ocfs2_dx_root_block *dx_root;
4084
4085         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4086         next_block = le64_to_cpu(dx_root->dr_free_blk);
4087
4088         while (next_block) {
4089                 brelse(prev_leaf_bh);
4090                 prev_leaf_bh = leaf_bh;
4091                 leaf_bh = NULL;
4092
4093                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4094                 if (ret) {
4095                         mlog_errno(ret);
4096                         goto out;
4097                 }
4098
4099                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4100                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4101                         lookup->dl_leaf_bh = leaf_bh;
4102                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
4103                         leaf_bh = NULL;
4104                         prev_leaf_bh = NULL;
4105                         break;
4106                 }
4107
4108                 next_block = le64_to_cpu(db->db_free_next);
4109         }
4110
4111         if (!next_block)
4112                 ret = -ENOSPC;
4113
4114 out:
4115
4116         brelse(leaf_bh);
4117         brelse(prev_leaf_bh);
4118         return ret;
4119 }
4120
4121 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4122                                        struct buffer_head *dx_root_bh)
4123 {
4124         int ret, num_dx_leaves, i, j, did_quota = 0;
4125         struct buffer_head **dx_leaves = NULL;
4126         struct ocfs2_extent_tree et;
4127         u64 insert_blkno;
4128         struct ocfs2_alloc_context *data_ac = NULL;
4129         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4130         handle_t *handle = NULL;
4131         struct ocfs2_dx_root_block *dx_root;
4132         struct ocfs2_dx_entry_list *entry_list;
4133         struct ocfs2_dx_entry *dx_entry;
4134         struct ocfs2_dx_leaf *target_leaf;
4135
4136         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4137         if (ret) {
4138                 mlog_errno(ret);
4139                 goto out;
4140         }
4141
4142         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4143         if (!dx_leaves) {
4144                 ret = -ENOMEM;
4145                 mlog_errno(ret);
4146                 goto out;
4147         }
4148
4149         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4150         if (IS_ERR(handle)) {
4151                 ret = PTR_ERR(handle);
4152                 mlog_errno(ret);
4153                 goto out;
4154         }
4155
4156         if (vfs_dq_alloc_space_nodirty(dir,
4157                                        ocfs2_clusters_to_bytes(osb->sb, 1))) {
4158                 ret = -EDQUOT;
4159                 goto out_commit;
4160         }
4161         did_quota = 1;
4162
4163         /*
4164          * We do this up front, before the allocation, so that a
4165          * failure to add the dx_root_bh to the journal won't result
4166          * us losing clusters.
4167          */
4168         ret = ocfs2_journal_access_dr(handle, dir, dx_root_bh,
4169                                       OCFS2_JOURNAL_ACCESS_WRITE);
4170         if (ret) {
4171                 mlog_errno(ret);
4172                 goto out_commit;
4173         }
4174
4175         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4176                                          num_dx_leaves, &insert_blkno);
4177         if (ret) {
4178                 mlog_errno(ret);
4179                 goto out_commit;
4180         }
4181
4182         /*
4183          * Transfer the entries from our dx_root into the appropriate
4184          * block
4185          */
4186         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4187         entry_list = &dx_root->dr_entries;
4188
4189         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4190                 dx_entry = &entry_list->de_entries[i];
4191
4192                 j = __ocfs2_dx_dir_hash_idx(osb,
4193                                             le32_to_cpu(dx_entry->dx_minor_hash));
4194                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4195
4196                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4197
4198                 /* Each leaf has been passed to the journal already
4199                  * via __ocfs2_dx_dir_new_cluster() */
4200         }
4201
4202         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4203         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4204                offsetof(struct ocfs2_dx_root_block, dr_list));
4205         dx_root->dr_list.l_count =
4206                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4207
4208         /* This should never fail considering we start with an empty
4209          * dx_root. */
4210         ocfs2_init_dx_root_extent_tree(&et, dir, dx_root_bh);
4211         ret = ocfs2_insert_extent(osb, handle, dir, &et, 0,
4212                                   insert_blkno, 1, 0, NULL);
4213         if (ret)
4214                 mlog_errno(ret);
4215         did_quota = 0;
4216
4217         ocfs2_journal_dirty(handle, dx_root_bh);
4218
4219 out_commit:
4220         if (ret < 0 && did_quota)
4221                 vfs_dq_free_space_nodirty(dir,
4222                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4223
4224         ocfs2_commit_trans(osb, handle);
4225
4226 out:
4227         if (data_ac)
4228                 ocfs2_free_alloc_context(data_ac);
4229
4230         if (dx_leaves) {
4231                 for (i = 0; i < num_dx_leaves; i++)
4232                         brelse(dx_leaves[i]);
4233                 kfree(dx_leaves);
4234         }
4235         return ret;
4236 }
4237
4238 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4239 {
4240         struct ocfs2_dx_root_block *dx_root;
4241         struct ocfs2_dx_entry_list *entry_list;
4242
4243         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4244         entry_list = &dx_root->dr_entries;
4245
4246         if (le16_to_cpu(entry_list->de_num_used) >=
4247             le16_to_cpu(entry_list->de_count))
4248                 return -ENOSPC;
4249
4250         return 0;
4251 }
4252
4253 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4254                                            struct buffer_head *di_bh,
4255                                            const char *name,
4256                                            int namelen,
4257                                            struct ocfs2_dir_lookup_result *lookup)
4258 {
4259         int ret, free_dx_root = 1;
4260         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4261         struct buffer_head *dx_root_bh = NULL;
4262         struct buffer_head *leaf_bh = NULL;
4263         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4264         struct ocfs2_dx_root_block *dx_root;
4265
4266         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4267         if (ret) {
4268                 mlog_errno(ret);
4269                 goto out;
4270         }
4271
4272         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4273         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4274                 ret = -ENOSPC;
4275                 mlog_errno(ret);
4276                 goto out;
4277         }
4278
4279         if (ocfs2_dx_root_inline(dx_root)) {
4280                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4281
4282                 if (ret == 0)
4283                         goto search_el;
4284
4285                 /*
4286                  * We ran out of room in the root block. Expand it to
4287                  * an extent, then allow ocfs2_find_dir_space_dx to do
4288                  * the rest.
4289                  */
4290                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4291                 if (ret) {
4292                         mlog_errno(ret);
4293                         goto out;
4294                 }
4295         }
4296
4297         /*
4298          * Insert preparation for an indexed directory is split into two
4299          * steps. The call to find_dir_space_dx reserves room in the index for
4300          * an additional item. If we run out of space there, it's a real error
4301          * we can't continue on.
4302          */
4303         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4304                                       namelen, lookup);
4305         if (ret) {
4306                 mlog_errno(ret);
4307                 goto out;
4308         }
4309
4310 search_el:
4311         /*
4312          * Next, we need to find space in the unindexed tree. This call
4313          * searches using the free space linked list. If the unindexed tree
4314          * lacks sufficient space, we'll expand it below. The expansion code
4315          * is smart enough to add any new blocks to the free space list.
4316          */
4317         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4318         if (ret && ret != -ENOSPC) {
4319                 mlog_errno(ret);
4320                 goto out;
4321         }
4322
4323         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4324         lookup->dl_dx_root_bh = dx_root_bh;
4325         free_dx_root = 0;
4326
4327         if (ret == -ENOSPC) {
4328                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4329
4330                 if (ret) {
4331                         mlog_errno(ret);
4332                         goto out;
4333                 }
4334
4335                 /*
4336                  * We make the assumption here that new leaf blocks are added
4337                  * to the front of our free list.
4338                  */
4339                 lookup->dl_prev_leaf_bh = NULL;
4340                 lookup->dl_leaf_bh = leaf_bh;
4341         }
4342
4343 out:
4344         if (free_dx_root)
4345                 brelse(dx_root_bh);
4346         return ret;
4347 }
4348
4349 /*
4350  * Get a directory ready for insert. Any directory allocation required
4351  * happens here. Success returns zero, and enough context in the dir
4352  * lookup result that ocfs2_add_entry() will be able complete the task
4353  * with minimal performance impact.
4354  */
4355 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4356                                  struct inode *dir,
4357                                  struct buffer_head *parent_fe_bh,
4358                                  const char *name,
4359                                  int namelen,
4360                                  struct ocfs2_dir_lookup_result *lookup)
4361 {
4362         int ret;
4363         unsigned int blocks_wanted = 1;
4364         struct buffer_head *bh = NULL;
4365
4366         mlog(0, "getting ready to insert namelen %d into dir %llu\n",
4367              namelen, (unsigned long long)OCFS2_I(dir)->ip_blkno);
4368
4369         if (!namelen) {
4370                 ret = -EINVAL;
4371                 mlog_errno(ret);
4372                 goto out;
4373         }
4374
4375         /*
4376          * Do this up front to reduce confusion.
4377          *
4378          * The directory might start inline, then be turned into an
4379          * indexed one, in which case we'd need to hash deep inside
4380          * ocfs2_find_dir_space_id(). Since
4381          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4382          * done, there seems no point in spreading out the calls. We
4383          * can optimize away the case where the file system doesn't
4384          * support indexing.
4385          */
4386         if (ocfs2_supports_indexed_dirs(osb))
4387                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4388
4389         if (ocfs2_dir_indexed(dir)) {
4390                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4391                                                       name, namelen, lookup);
4392                 if (ret)
4393                         mlog_errno(ret);
4394                 goto out;
4395         }
4396
4397         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4398                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4399                                               namelen, &bh, &blocks_wanted);
4400         } else
4401                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4402
4403         if (ret && ret != -ENOSPC) {
4404                 mlog_errno(ret);
4405                 goto out;
4406         }
4407
4408         if (ret == -ENOSPC) {
4409                 /*
4410                  * We have to expand the directory to add this name.
4411                  */
4412                 BUG_ON(bh);
4413
4414                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4415                                        lookup, &bh);
4416                 if (ret) {
4417                         if (ret != -ENOSPC)
4418                                 mlog_errno(ret);
4419                         goto out;
4420                 }
4421
4422                 BUG_ON(!bh);
4423         }
4424
4425         lookup->dl_leaf_bh = bh;
4426         bh = NULL;
4427 out:
4428         brelse(bh);
4429         return ret;
4430 }
4431
4432 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4433                                      struct buffer_head *di_bh,
4434                                      struct buffer_head *dx_root_bh)
4435 {
4436         int ret;
4437         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4438         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4439         struct ocfs2_dx_root_block *dx_root;
4440         struct inode *dx_alloc_inode = NULL;
4441         struct buffer_head *dx_alloc_bh = NULL;
4442         handle_t *handle;
4443         u64 blk;
4444         u16 bit;
4445         u64 bg_blkno;
4446
4447         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4448
4449         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4450                                         EXTENT_ALLOC_SYSTEM_INODE,
4451                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4452         if (!dx_alloc_inode) {
4453                 ret = -ENOMEM;
4454                 mlog_errno(ret);
4455                 goto out;
4456         }
4457         mutex_lock(&dx_alloc_inode->i_mutex);
4458
4459         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4460         if (ret) {
4461                 mlog_errno(ret);
4462                 goto out_mutex;
4463         }
4464
4465         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4466         if (IS_ERR(handle)) {
4467                 ret = PTR_ERR(handle);
4468                 mlog_errno(ret);
4469                 goto out_unlock;
4470         }
4471
4472         ret = ocfs2_journal_access_di(handle, dir, di_bh,
4473                                       OCFS2_JOURNAL_ACCESS_WRITE);
4474         if (ret) {
4475                 mlog_errno(ret);
4476                 goto out_commit;
4477         }
4478
4479         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4480         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4481         di->i_dx_root = cpu_to_le64(0ULL);
4482
4483         ocfs2_journal_dirty(handle, di_bh);
4484
4485         blk = le64_to_cpu(dx_root->dr_blkno);
4486         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4487         bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4488         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4489                                        bit, bg_blkno, 1);
4490         if (ret)
4491                 mlog_errno(ret);
4492
4493 out_commit:
4494         ocfs2_commit_trans(osb, handle);
4495
4496 out_unlock:
4497         ocfs2_inode_unlock(dx_alloc_inode, 1);
4498
4499 out_mutex:
4500         mutex_unlock(&dx_alloc_inode->i_mutex);
4501         brelse(dx_alloc_bh);
4502 out:
4503         iput(dx_alloc_inode);
4504         return ret;
4505 }
4506
4507 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4508 {
4509         int ret;
4510         unsigned int uninitialized_var(clen);
4511         u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4512         u64 uninitialized_var(blkno);
4513         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4514         struct buffer_head *dx_root_bh = NULL;
4515         struct ocfs2_dx_root_block *dx_root;
4516         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4517         struct ocfs2_cached_dealloc_ctxt dealloc;
4518         struct ocfs2_extent_tree et;
4519
4520         ocfs2_init_dealloc_ctxt(&dealloc);
4521
4522         if (!ocfs2_dir_indexed(dir))
4523                 return 0;
4524
4525         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4526         if (ret) {
4527                 mlog_errno(ret);
4528                 goto out;
4529         }
4530         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4531
4532         if (ocfs2_dx_root_inline(dx_root))
4533                 goto remove_index;
4534
4535         ocfs2_init_dx_root_extent_tree(&et, dir, dx_root_bh);
4536
4537         /* XXX: What if dr_clusters is too large? */
4538         while (le32_to_cpu(dx_root->dr_clusters)) {
4539                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4540                                               major_hash, &cpos, &blkno, &clen);
4541                 if (ret) {
4542                         mlog_errno(ret);
4543                         goto out;
4544                 }
4545
4546                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4547
4548                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen,
4549                                                &dealloc);
4550                 if (ret) {
4551                         mlog_errno(ret);
4552                         goto out;
4553                 }
4554
4555                 if (cpos == 0)
4556                         break;
4557
4558                 major_hash = cpos - 1;
4559         }
4560
4561 remove_index:
4562         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4563         if (ret) {
4564                 mlog_errno(ret);
4565                 goto out;
4566         }
4567
4568         ocfs2_remove_from_cache(dir, dx_root_bh);
4569 out:
4570         ocfs2_schedule_truncate_log_flush(osb, 1);
4571         ocfs2_run_deallocs(osb, &dealloc);
4572
4573         brelse(dx_root_bh);
4574         return ret;
4575 }