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