fs/ocfs2/dir.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dir.c
   5  *
   6  * Creates, reads, walks and deletes directory-nodes
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  *  Portions of this code from linux/fs/ext3/dir.c
  11  *
  12  *  Copyright (C) 1992, 1993, 1994, 1995
  13  *  Remy Card (card@masi.ibp.fr)
  14  *  Laboratoire MASI - Institut Blaise pascal
  15  *  Universite Pierre et Marie Curie (Paris VI)
  16  *
  17  *   from
  18  *
  19  *   linux/fs/minix/dir.c
  20  *
  21  *   Copyright (C) 1991, 1992 Linux Torvalds
  22  *
  23  * This program is free software; you can redistribute it and/or
  24  * modify it under the terms of the GNU General Public
  25  * License as published by the Free Software Foundation; either
  26  * version 2 of the License, or (at your option) any later version.
  27  *
  28  * This program is distributed in the hope that it will be useful,
  29  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  30  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  31  * General Public License for more details.
  32  *
  33  * You should have received a copy of the GNU General Public
  34  * License along with this program; if not, write to the
  35  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  36  * Boston, MA 021110-1307, USA.
  37  */
  38
  39 #include <linux/fs.h>
  40 #include <linux/types.h>
  41 #include <linux/slab.h>
  42 #include <linux/highmem.h>
  43 #include <linux/quotaops.h>
  44 #include <linux/sort.h>
  45
  46 #include <cluster/masklog.h>
  47
  48 #include "ocfs2.h"
  49
  50 #include "alloc.h"
  51 #include "blockcheck.h"
  52 #include "dir.h"
  53 #include "dlmglue.h"
  54 #include "extent_map.h"
  55 #include "file.h"
  56 #include "inode.h"
  57 #include "journal.h"
  58 #include "namei.h"
  59 #include "suballoc.h"
  60 #include "super.h"
  61 #include "sysfile.h"
  62 #include "uptodate.h"
  63 #include "ocfs2_trace.h"
  64
  65 #include "buffer_head_io.h"
  66
  67 #define NAMEI_RA_CHUNKS  2
  68 #define NAMEI_RA_BLOCKS  4
  69 #define NAMEI_RA_SIZE        (NAMEI_RA_CHUNKS * NAMEI_RA_BLOCKS)
  70
  71 static unsigned char ocfs2_filetype_table[] = {
  72         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
  73 };
  74
  75 static int ocfs2_do_extend_dir(struct super_block *sb,
  76                                handle_t *handle,
  77                                struct inode *dir,
  78                                struct buffer_head *parent_fe_bh,
  79                                struct ocfs2_alloc_context *data_ac,
  80                                struct ocfs2_alloc_context *meta_ac,
  81                                struct buffer_head **new_bh);
  82 static int ocfs2_dir_indexed(struct inode *inode);
  83
  84 /*
  85  * These are distinct checks because future versions of the file system will
  86  * want to have a trailing dirent structure independent of indexing.
  87  */
  88 static int ocfs2_supports_dir_trailer(struct inode *dir)
  89 {
  90         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
  91
  92         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
  93                 return 0;
  94
  95         return ocfs2_meta_ecc(osb) || ocfs2_dir_indexed(dir);
  96 }
  97
  98 /*
  99  * "new' here refers to the point at which we're creating a new
 100  * directory via "mkdir()", but also when we're expanding an inline
 101  * directory. In either case, we don't yet have the indexing bit set
 102  * on the directory, so the standard checks will fail in when metaecc
 103  * is turned off. Only directory-initialization type functions should
 104  * use this then. Everything else wants ocfs2_supports_dir_trailer()
 105  */
 106 static int ocfs2_new_dir_wants_trailer(struct inode *dir)
 107 {
 108         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
 109
 110         return ocfs2_meta_ecc(osb) ||
 111                 ocfs2_supports_indexed_dirs(osb);
 112 }
 113
 114 static inline unsigned int ocfs2_dir_trailer_blk_off(struct super_block *sb)
 115 {
 116         return sb->s_blocksize - sizeof(struct ocfs2_dir_block_trailer);
 117 }
 118
 119 #define ocfs2_trailer_from_bh(_bh, _sb) ((struct ocfs2_dir_block_trailer *) ((_bh)->b_data + ocfs2_dir_trailer_blk_off((_sb))))
 120
 121 /* XXX ocfs2_block_dqtrailer() is similar but not quite - can we make
 122  * them more consistent? */
 123 struct ocfs2_dir_block_trailer *ocfs2_dir_trailer_from_size(int blocksize,
 124                                                             void *data)
 125 {
 126         char *p = data;
 127
 128         p += blocksize - sizeof(struct ocfs2_dir_block_trailer);
 129         return (struct ocfs2_dir_block_trailer *)p;
 130 }
 131
 132 /*
 133  * XXX: This is executed once on every dirent. We should consider optimizing
 134  * it.
 135  */
 136 static int ocfs2_skip_dir_trailer(struct inode *dir,
 137                                   struct ocfs2_dir_entry *de,
 138                                   unsigned long offset,
 139                                   unsigned long blklen)
 140 {
 141         unsigned long toff = blklen - sizeof(struct ocfs2_dir_block_trailer);
 142
 143         if (!ocfs2_supports_dir_trailer(dir))
 144                 return 0;
 145
 146         if (offset != toff)
 147                 return 0;
 148
 149         return 1;
 150 }
 151
 152 static void ocfs2_init_dir_trailer(struct inode *inode,
 153                                    struct buffer_head *bh, u16 rec_len)
 154 {
 155         struct ocfs2_dir_block_trailer *trailer;
 156
 157         trailer = ocfs2_trailer_from_bh(bh, inode->i_sb);
 158         strcpy(trailer->db_signature, OCFS2_DIR_TRAILER_SIGNATURE);
 159         trailer->db_compat_rec_len =
 160                         cpu_to_le16(sizeof(struct ocfs2_dir_block_trailer));
 161         trailer->db_parent_dinode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
 162         trailer->db_blkno = cpu_to_le64(bh->b_blocknr);
 163         trailer->db_free_rec_len = cpu_to_le16(rec_len);
 164 }
 165 /*
 166  * Link an unindexed block with a dir trailer structure into the index free
 167  * list. This function will modify dirdata_bh, but assumes you've already
 168  * passed it to the journal.
 169  */
 170 static int ocfs2_dx_dir_link_trailer(struct inode *dir, handle_t *handle,
 171                                      struct buffer_head *dx_root_bh,
 172                                      struct buffer_head *dirdata_bh)
 173 {
 174         int ret;
 175         struct ocfs2_dx_root_block *dx_root;
 176         struct ocfs2_dir_block_trailer *trailer;
 177
 178         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
 179                                       OCFS2_JOURNAL_ACCESS_WRITE);
 180         if (ret) {
 181                 mlog_errno(ret);
 182                 goto out;
 183         }
 184         trailer = ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
 185         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
 186
 187         trailer->db_free_next = dx_root->dr_free_blk;
 188         dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
 189
 190         ocfs2_journal_dirty(handle, dx_root_bh);
 191
 192 out:
 193         return ret;
 194 }
 195
 196 static int ocfs2_free_list_at_root(struct ocfs2_dir_lookup_result *res)
 197 {
 198         return res->dl_prev_leaf_bh == NULL;
 199 }
 200
 201 void ocfs2_free_dir_lookup_result(struct ocfs2_dir_lookup_result *res)
 202 {
 203         brelse(res->dl_dx_root_bh);
 204         brelse(res->dl_leaf_bh);
 205         brelse(res->dl_dx_leaf_bh);
 206         brelse(res->dl_prev_leaf_bh);
 207 }
 208
 209 static int ocfs2_dir_indexed(struct inode *inode)
 210 {
 211         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INDEXED_DIR_FL)
 212                 return 1;
 213         return 0;
 214 }
 215
 216 static inline int ocfs2_dx_root_inline(struct ocfs2_dx_root_block *dx_root)
 217 {
 218         return dx_root->dr_flags & OCFS2_DX_FLAG_INLINE;
 219 }
 220
 221 /*
 222  * Hashing code adapted from ext3
 223  */
 224 #define DELTA 0x9E3779B9
 225
 226 static void TEA_transform(__u32 buf[4], __u32 const in[])
 227 {
 228         __u32   sum = 0;
 229         __u32   b0 = buf[0], b1 = buf[1];
 230         __u32   a = in[0], b = in[1], c = in[2], d = in[3];
 231         int     n = 16;
 232
 233         do {
 234                 sum += DELTA;
 235                 b0 += ((b1 << 4)+a) ^ (b1+sum) ^ ((b1 >> 5)+b);
 236                 b1 += ((b0 << 4)+c) ^ (b0+sum) ^ ((b0 >> 5)+d);
 237         } while (--n);
 238
 239         buf[0] += b0;
 240         buf[1] += b1;
 241 }
 242
 243 static void str2hashbuf(const char *msg, int len, __u32 *buf, int num)
 244 {
 245         __u32   pad, val;
 246         int     i;
 247
 248         pad = (__u32)len | ((__u32)len << 8);
 249         pad |= pad << 16;
 250
 251         val = pad;
 252         if (len > num*4)
 253                 len = num * 4;
 254         for (i = 0; i < len; i++) {
 255                 if ((i % 4) == 0)
 256                         val = pad;
 257                 val = msg[i] + (val << 8);
 258                 if ((i % 4) == 3) {
 259                         *buf++ = val;
 260                         val = pad;
 261                         num--;
 262                 }
 263         }
 264         if (--num >= 0)
 265                 *buf++ = val;
 266         while (--num >= 0)
 267                 *buf++ = pad;
 268 }
 269
 270 static void ocfs2_dx_dir_name_hash(struct inode *dir, const char *name, int len,
 271                                    struct ocfs2_dx_hinfo *hinfo)
 272 {
 273         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
 274         const char      *p;
 275         __u32           in[8], buf[4];
 276
 277         /*
 278          * XXX: Is this really necessary, if the index is never looked
 279          * at by readdir? Is a hash value of '0' a bad idea?
 280          */
 281         if ((len == 1 && !strncmp(".", name, 1)) ||
 282             (len == 2 && !strncmp("..", name, 2))) {
 283                 buf[0] = buf[1] = 0;
 284                 goto out;
 285         }
 286
 287 #ifdef OCFS2_DEBUG_DX_DIRS
 288         /*
 289          * This makes it very easy to debug indexing problems. We
 290          * should never allow this to be selected without hand editing
 291          * this file though.
 292          */
 293         buf[0] = buf[1] = len;
 294         goto out;
 295 #endif
 296
 297         memcpy(buf, osb->osb_dx_seed, sizeof(buf));
 298
 299         p = name;
 300         while (len > 0) {
 301                 str2hashbuf(p, len, in, 4);
 302                 TEA_transform(buf, in);
 303                 len -= 16;
 304                 p += 16;
 305         }
 306
 307 out:
 308         hinfo->major_hash = buf[0];
 309         hinfo->minor_hash = buf[1];
 310 }
 311
 312 /*
 313  * bh passed here can be an inode block or a dir data block, depending
 314  * on the inode inline data flag.
 315  */
 316 static int ocfs2_check_dir_entry(struct inode * dir,
 317                                  struct ocfs2_dir_entry * de,
 318                                  struct buffer_head * bh,
 319                                  unsigned long offset)
 320 {
 321         const char *error_msg = NULL;
 322         const int rlen = le16_to_cpu(de->rec_len);
 323
 324         if (unlikely(rlen < OCFS2_DIR_REC_LEN(1)))
 325                 error_msg = "rec_len is smaller than minimal";
 326         else if (unlikely(rlen % 4 != 0))
 327                 error_msg = "rec_len % 4 != 0";
 328         else if (unlikely(rlen < OCFS2_DIR_REC_LEN(de->name_len)))
 329                 error_msg = "rec_len is too small for name_len";
 330         else if (unlikely(
 331                  ((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize))
 332                 error_msg = "directory entry across blocks";
 333
 334         if (unlikely(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
 341         return error_msg == NULL ? 1 : 0;
 342 }
 343
 344 static inline int ocfs2_match(int len,
 345                               const char * const name,
 346                               struct ocfs2_dir_entry *de)
 347 {
 348         if (len != de->name_len)
 349                 return 0;
 350         if (!de->inode)
 351                 return 0;
 352         return !memcmp(name, de->name, len);
 353 }
 354
 355 /*
 356  * Returns 0 if not found, -1 on failure, and 1 on success
 357  */
 358 static inline int ocfs2_search_dirblock(struct buffer_head *bh,
 359                                         struct inode *dir,
 360                                         const char *name, int namelen,
 361                                         unsigned long offset,
 362                                         char *first_de,
 363                                         unsigned int bytes,
 364                                         struct ocfs2_dir_entry **res_dir)
 365 {
 366         struct ocfs2_dir_entry *de;
 367         char *dlimit, *de_buf;
 368         int de_len;
 369         int ret = 0;
 370
 371         de_buf = first_de;
 372         dlimit = de_buf + bytes;
 373
 374         while (de_buf < dlimit) {
 375                 /* this code is executed quadratically often */
 376                 /* do minimal checking `by hand' */
 377
 378                 de = (struct ocfs2_dir_entry *) de_buf;
 379
 380                 if (de_buf + namelen <= dlimit &&
 381                     ocfs2_match(namelen, name, de)) {
 382                         /* found a match - just to be sure, do a full check */
 383                         if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
 384                                 ret = -1;
 385                                 goto bail;
 386                         }
 387                         *res_dir = de;
 388                         ret = 1;
 389                         goto bail;
 390                 }
 391
 392                 /* prevent looping on a bad block */
 393                 de_len = le16_to_cpu(de->rec_len);
 394                 if (de_len <= 0) {
 395                         ret = -1;
 396                         goto bail;
 397                 }
 398
 399                 de_buf += de_len;
 400                 offset += de_len;
 401         }
 402
 403 bail:
 404         trace_ocfs2_search_dirblock(ret);
 405         return ret;
 406 }
 407
 408 static struct buffer_head *ocfs2_find_entry_id(const char *name,
 409                                                int namelen,
 410                                                struct inode *dir,
 411                                                struct ocfs2_dir_entry **res_dir)
 412 {
 413         int ret, found;
 414         struct buffer_head *di_bh = NULL;
 415         struct ocfs2_dinode *di;
 416         struct ocfs2_inline_data *data;
 417
 418         ret = ocfs2_read_inode_block(dir, &di_bh);
 419         if (ret) {
 420                 mlog_errno(ret);
 421                 goto out;
 422         }
 423
 424         di = (struct ocfs2_dinode *)di_bh->b_data;
 425         data = &di->id2.i_data;
 426
 427         found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
 428                                       data->id_data, i_size_read(dir), res_dir);
 429         if (found == 1)
 430                 return di_bh;
 431
 432         brelse(di_bh);
 433 out:
 434         return NULL;
 435 }
 436
 437 static int ocfs2_validate_dir_block(struct super_block *sb,
 438                                     struct buffer_head *bh)
 439 {
 440         int rc;
 441         struct ocfs2_dir_block_trailer *trailer =
 442                 ocfs2_trailer_from_bh(bh, sb);
 443
 444
 445         /*
 446          * We don't validate dirents here, that's handled
 447          * in-place when the code walks them.
 448          */
 449         trace_ocfs2_validate_dir_block((unsigned long long)bh->b_blocknr);
 450
 451         BUG_ON(!buffer_uptodate(bh));
 452
 453         /*
 454          * If the ecc fails, we return the error but otherwise
 455          * leave the filesystem running.  We know any error is
 456          * local to this block.
 457          *
 458          * Note that we are safe to call this even if the directory
 459          * doesn't have a trailer.  Filesystems without metaecc will do
 460          * nothing, and filesystems with it will have one.
 461          */
 462         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
 463         if (rc)
 464                 mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
 465                      (unsigned long long)bh->b_blocknr);
 466
 467         return rc;
 468 }
 469
 470 /*
 471  * Validate a directory trailer.
 472  *
 473  * We check the trailer here rather than in ocfs2_validate_dir_block()
 474  * because that function doesn't have the inode to test.
 475  */
 476 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
 477 {
 478         int rc = 0;
 479         struct ocfs2_dir_block_trailer *trailer;
 480
 481         trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
 482         if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
 483                 rc = -EINVAL;
 484                 ocfs2_error(dir->i_sb,
 485                             "Invalid dirblock #%llu: "
 486                             "signature = %.*s\n",
 487                             (unsigned long long)bh->b_blocknr, 7,
 488                             trailer->db_signature);
 489                 goto out;
 490         }
 491         if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
 492                 rc = -EINVAL;
 493                 ocfs2_error(dir->i_sb,
 494                             "Directory block #%llu has an invalid "
 495                             "db_blkno of %llu",
 496                             (unsigned long long)bh->b_blocknr,
 497                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
 498                 goto out;
 499         }
 500         if (le64_to_cpu(trailer->db_parent_dinode) !=
 501             OCFS2_I(dir)->ip_blkno) {
 502                 rc = -EINVAL;
 503                 ocfs2_error(dir->i_sb,
 504                             "Directory block #%llu on dinode "
 505                             "#%llu has an invalid parent_dinode "
 506                             "of %llu",
 507                             (unsigned long long)bh->b_blocknr,
 508                             (unsigned long long)OCFS2_I(dir)->ip_blkno,
 509                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
 510                 goto out;
 511         }
 512 out:
 513         return rc;
 514 }
 515
 516 /*
 517  * This function forces all errors to -EIO for consistency with its
 518  * predecessor, ocfs2_bread().  We haven't audited what returning the
 519  * real error codes would do to callers.  We log the real codes with
 520  * mlog_errno() before we squash them.
 521  */
 522 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
 523                                 struct buffer_head **bh, int flags)
 524 {
 525         int rc = 0;
 526         struct buffer_head *tmp = *bh;
 527
 528         rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
 529                                     ocfs2_validate_dir_block);
 530         if (rc) {
 531                 mlog_errno(rc);
 532                 goto out;
 533         }
 534
 535         if (!(flags & OCFS2_BH_READAHEAD) &&
 536             ocfs2_supports_dir_trailer(inode)) {
 537                 rc = ocfs2_check_dir_trailer(inode, tmp);
 538                 if (rc) {
 539                         if (!*bh)
 540                                 brelse(tmp);
 541                         mlog_errno(rc);
 542                         goto out;
 543                 }
 544         }
 545
 546         /* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
 547         if (!*bh)
 548                 *bh = tmp;
 549
 550 out:
 551         return rc ? -EIO : 0;
 552 }
 553
 554 /*
 555  * Read the block at 'phys' which belongs to this directory
 556  * inode. This function does no virtual->physical block translation -
 557  * what's passed in is assumed to be a valid directory block.
 558  */
 559 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
 560                                        struct buffer_head **bh)
 561 {
 562         int ret;
 563         struct buffer_head *tmp = *bh;
 564
 565         ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
 566                                ocfs2_validate_dir_block);
 567         if (ret) {
 568                 mlog_errno(ret);
 569                 goto out;
 570         }
 571
 572         if (ocfs2_supports_dir_trailer(dir)) {
 573                 ret = ocfs2_check_dir_trailer(dir, tmp);
 574                 if (ret) {
 575                         if (!*bh)
 576                                 brelse(tmp);
 577                         mlog_errno(ret);
 578                         goto out;
 579                 }
 580         }
 581
 582         if (!ret && !*bh)
 583                 *bh = tmp;
 584 out:
 585         return ret;
 586 }
 587
 588 static int ocfs2_validate_dx_root(struct super_block *sb,
 589                                   struct buffer_head *bh)
 590 {
 591         int ret;
 592         struct ocfs2_dx_root_block *dx_root;
 593
 594         BUG_ON(!buffer_uptodate(bh));
 595
 596         dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
 597
 598         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
 599         if (ret) {
 600                 mlog(ML_ERROR,
 601                      "Checksum failed for dir index root block %llu\n",
 602                      (unsigned long long)bh->b_blocknr);
 603                 return ret;
 604         }
 605
 606         if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
 607                 ocfs2_error(sb,
 608                             "Dir Index Root # %llu has bad signature %.*s",
 609                             (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
 610                             7, dx_root->dr_signature);
 611                 return -EINVAL;
 612         }
 613
 614         return 0;
 615 }
 616
 617 static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
 618                               struct buffer_head **dx_root_bh)
 619 {
 620         int ret;
 621         u64 blkno = le64_to_cpu(di->i_dx_root);
 622         struct buffer_head *tmp = *dx_root_bh;
 623
 624         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 625                                ocfs2_validate_dx_root);
 626
 627         /* If ocfs2_read_block() got us a new bh, pass it up. */
 628         if (!ret && !*dx_root_bh)
 629                 *dx_root_bh = tmp;
 630
 631         return ret;
 632 }
 633
 634 static int ocfs2_validate_dx_leaf(struct super_block *sb,
 635                                   struct buffer_head *bh)
 636 {
 637         int ret;
 638         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
 639
 640         BUG_ON(!buffer_uptodate(bh));
 641
 642         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
 643         if (ret) {
 644                 mlog(ML_ERROR,
 645                      "Checksum failed for dir index leaf block %llu\n",
 646                      (unsigned long long)bh->b_blocknr);
 647                 return ret;
 648         }
 649
 650         if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
 651                 ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s",
 652                             7, dx_leaf->dl_signature);
 653                 return -EROFS;
 654         }
 655
 656         return 0;
 657 }
 658
 659 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
 660                               struct buffer_head **dx_leaf_bh)
 661 {
 662         int ret;
 663         struct buffer_head *tmp = *dx_leaf_bh;
 664
 665         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 666                                ocfs2_validate_dx_leaf);
 667
 668         /* If ocfs2_read_block() got us a new bh, pass it up. */
 669         if (!ret && !*dx_leaf_bh)
 670                 *dx_leaf_bh = tmp;
 671
 672         return ret;
 673 }
 674
 675 /*
 676  * Read a series of dx_leaf blocks. This expects all buffer_head
 677  * pointers to be NULL on function entry.
 678  */
 679 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
 680                                 struct buffer_head **dx_leaf_bhs)
 681 {
 682         int ret;
 683
 684         ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
 685                                 ocfs2_validate_dx_leaf);
 686         if (ret)
 687                 mlog_errno(ret);
 688
 689         return ret;
 690 }
 691
 692 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
 693                                                struct inode *dir,
 694                                                struct ocfs2_dir_entry **res_dir)
 695 {
 696         struct super_block *sb;
 697         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 698         struct buffer_head *bh, *ret = NULL;
 699         unsigned long start, block, b;
 700         int ra_max = 0;         /* Number of bh's in the readahead
 701                                    buffer, bh_use[] */
 702         int ra_ptr = 0;         /* Current index into readahead
 703                                    buffer */
 704         int num = 0;
 705         int nblocks, i, err;
 706
 707         sb = dir->i_sb;
 708
 709         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 710         start = OCFS2_I(dir)->ip_dir_start_lookup;
 711         if (start >= nblocks)
 712                 start = 0;
 713         block = start;
 714
 715 restart:
 716         do {
 717                 /*
 718                  * We deal with the read-ahead logic here.
 719                  */
 720                 if (ra_ptr >= ra_max) {
 721                         /* Refill the readahead buffer */
 722                         ra_ptr = 0;
 723                         b = block;
 724                         for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
 725                                 /*
 726                                  * Terminate if we reach the end of the
 727                                  * directory and must wrap, or if our
 728                                  * search has finished at this block.
 729                                  */
 730                                 if (b >= nblocks || (num && block == start)) {
 731                                         bh_use[ra_max] = NULL;
 732                                         break;
 733                                 }
 734                                 num++;
 735
 736                                 bh = NULL;
 737                                 err = ocfs2_read_dir_block(dir, b++, &bh,
 738                                                            OCFS2_BH_READAHEAD);
 739                                 bh_use[ra_max] = bh;
 740                         }
 741                 }
 742                 if ((bh = bh_use[ra_ptr++]) == NULL)
 743                         goto next;
 744                 if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
 745                         /* read error, skip block & hope for the best.
 746                          * ocfs2_read_dir_block() has released the bh. */
 747                         mlog(ML_ERROR, "reading directory %llu, "
 748                                     "offset %lu\n",
 749                                     (unsigned long long)OCFS2_I(dir)->ip_blkno,
 750                                     block);
 751                         goto next;
 752                 }
 753                 i = ocfs2_search_dirblock(bh, dir, name, namelen,
 754                                           block << sb->s_blocksize_bits,
 755                                           bh->b_data, sb->s_blocksize,
 756                                           res_dir);
 757                 if (i == 1) {
 758                         OCFS2_I(dir)->ip_dir_start_lookup = block;
 759                         ret = bh;
 760                         goto cleanup_and_exit;
 761                 } else {
 762                         brelse(bh);
 763                         if (i < 0)
 764                                 goto cleanup_and_exit;
 765                 }
 766         next:
 767                 if (++block >= nblocks)
 768                         block = 0;
 769         } while (block != start);
 770
 771         /*
 772          * If the directory has grown while we were searching, then
 773          * search the last part of the directory before giving up.
 774          */
 775         block = nblocks;
 776         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 777         if (block < nblocks) {
 778                 start = 0;
 779                 goto restart;
 780         }
 781
 782 cleanup_and_exit:
 783         /* Clean up the read-ahead blocks */
 784         for (; ra_ptr < ra_max; ra_ptr++)
 785                 brelse(bh_use[ra_ptr]);
 786
 787         trace_ocfs2_find_entry_el(ret);
 788         return ret;
 789 }
 790
 791 static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
 792                                    struct ocfs2_extent_list *el,
 793                                    u32 major_hash,
 794                                    u32 *ret_cpos,
 795                                    u64 *ret_phys_blkno,
 796                                    unsigned int *ret_clen)
 797 {
 798         int ret = 0, i, found;
 799         struct buffer_head *eb_bh = NULL;
 800         struct ocfs2_extent_block *eb;
 801         struct ocfs2_extent_rec *rec = NULL;
 802
 803         if (el->l_tree_depth) {
 804                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
 805                                       &eb_bh);
 806                 if (ret) {
 807                         mlog_errno(ret);
 808                         goto out;
 809                 }
 810
 811                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 812                 el = &eb->h_list;
 813
 814                 if (el->l_tree_depth) {
 815                         ocfs2_error(inode->i_sb,
 816                                     "Inode %lu has non zero tree depth in "
 817                                     "btree tree block %llu\n", inode->i_ino,
 818                                     (unsigned long long)eb_bh->b_blocknr);
 819                         ret = -EROFS;
 820                         goto out;
 821                 }
 822         }
 823
 824         found = 0;
 825         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
 826                 rec = &el->l_recs[i];
 827
 828                 if (le32_to_cpu(rec->e_cpos) <= major_hash) {
 829                         found = 1;
 830                         break;
 831                 }
 832         }
 833
 834         if (!found) {
 835                 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
 836                             "record (%u, %u, 0) in btree", inode->i_ino,
 837                             le32_to_cpu(rec->e_cpos),
 838                             ocfs2_rec_clusters(el, rec));
 839                 ret = -EROFS;
 840                 goto out;
 841         }
 842
 843         if (ret_phys_blkno)
 844                 *ret_phys_blkno = le64_to_cpu(rec->e_blkno);
 845         if (ret_cpos)
 846                 *ret_cpos = le32_to_cpu(rec->e_cpos);
 847         if (ret_clen)
 848                 *ret_clen = le16_to_cpu(rec->e_leaf_clusters);
 849
 850 out:
 851         brelse(eb_bh);
 852         return ret;
 853 }
 854
 855 /*
 856  * Returns the block index, from the start of the cluster which this
 857  * hash belongs too.
 858  */
 859 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 860                                                    u32 minor_hash)
 861 {
 862         return minor_hash & osb->osb_dx_mask;
 863 }
 864
 865 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 866                                           struct ocfs2_dx_hinfo *hinfo)
 867 {
 868         return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
 869 }
 870
 871 static int ocfs2_dx_dir_lookup(struct inode *inode,
 872                                struct ocfs2_extent_list *el,
 873                                struct ocfs2_dx_hinfo *hinfo,
 874                                u32 *ret_cpos,
 875                                u64 *ret_phys_blkno)
 876 {
 877         int ret = 0;
 878         unsigned int cend, uninitialized_var(clen);
 879         u32 uninitialized_var(cpos);
 880         u64 uninitialized_var(blkno);
 881         u32 name_hash = hinfo->major_hash;
 882
 883         ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
 884                                       &clen);
 885         if (ret) {
 886                 mlog_errno(ret);
 887                 goto out;
 888         }
 889
 890         cend = cpos + clen;
 891         if (name_hash >= cend) {
 892                 /* We want the last cluster */
 893                 blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
 894                 cpos += clen - 1;
 895         } else {
 896                 blkno += ocfs2_clusters_to_blocks(inode->i_sb,
 897                                                   name_hash - cpos);
 898                 cpos = name_hash;
 899         }
 900
 901         /*
 902          * We now have the cluster which should hold our entry. To
 903          * find the exact block from the start of the cluster to
 904          * search, we take the lower bits of the hash.
 905          */
 906         blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
 907
 908         if (ret_phys_blkno)
 909                 *ret_phys_blkno = blkno;
 910         if (ret_cpos)
 911                 *ret_cpos = cpos;
 912
 913 out:
 914
 915         return ret;
 916 }
 917
 918 static int ocfs2_dx_dir_search(const char *name, int namelen,
 919                                struct inode *dir,
 920                                struct ocfs2_dx_root_block *dx_root,
 921                                struct ocfs2_dir_lookup_result *res)
 922 {
 923         int ret, i, found;
 924         u64 uninitialized_var(phys);
 925         struct buffer_head *dx_leaf_bh = NULL;
 926         struct ocfs2_dx_leaf *dx_leaf;
 927         struct ocfs2_dx_entry *dx_entry = NULL;
 928         struct buffer_head *dir_ent_bh = NULL;
 929         struct ocfs2_dir_entry *dir_ent = NULL;
 930         struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
 931         struct ocfs2_extent_list *dr_el;
 932         struct ocfs2_dx_entry_list *entry_list;
 933
 934         ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
 935
 936         if (ocfs2_dx_root_inline(dx_root)) {
 937                 entry_list = &dx_root->dr_entries;
 938                 goto search;
 939         }
 940
 941         dr_el = &dx_root->dr_list;
 942
 943         ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
 944         if (ret) {
 945                 mlog_errno(ret);
 946                 goto out;
 947         }
 948
 949         trace_ocfs2_dx_dir_search((unsigned long long)OCFS2_I(dir)->ip_blkno,
 950                                   namelen, name, hinfo->major_hash,
 951                                   hinfo->minor_hash, (unsigned long long)phys);
 952
 953         ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
 954         if (ret) {
 955                 mlog_errno(ret);
 956                 goto out;
 957         }
 958
 959         dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
 960
 961         trace_ocfs2_dx_dir_search_leaf_info(
 962                         le16_to_cpu(dx_leaf->dl_list.de_num_used),
 963                         le16_to_cpu(dx_leaf->dl_list.de_count));
 964
 965         entry_list = &dx_leaf->dl_list;
 966
 967 search:
 968         /*
 969          * Empty leaf is legal, so no need to check for that.
 970          */
 971         found = 0;
 972         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
 973                 dx_entry = &entry_list->de_entries[i];
 974
 975                 if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
 976                     || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
 977                         continue;
 978
 979                 /*
 980                  * Search unindexed leaf block now. We're not
 981                  * guaranteed to find anything.
 982                  */
 983                 ret = ocfs2_read_dir_block_direct(dir,
 984                                           le64_to_cpu(dx_entry->dx_dirent_blk),
 985                                           &dir_ent_bh);
 986                 if (ret) {
 987                         mlog_errno(ret);
 988                         goto out;
 989                 }
 990
 991                 /*
 992                  * XXX: We should check the unindexed block here,
 993                  * before using it.
 994                  */
 995
 996                 found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
 997                                               0, dir_ent_bh->b_data,
 998                                               dir->i_sb->s_blocksize, &dir_ent);
 999                 if (found == 1)
1000                         break;
1001
1002                 if (found == -1) {
1003                         /* This means we found a bad directory entry. */
1004                         ret = -EIO;
1005                         mlog_errno(ret);
1006                         goto out;
1007                 }
1008
1009                 brelse(dir_ent_bh);
1010                 dir_ent_bh = NULL;
1011         }
1012
1013         if (found <= 0) {
1014                 ret = -ENOENT;
1015                 goto out;
1016         }
1017
1018         res->dl_leaf_bh = dir_ent_bh;
1019         res->dl_entry = dir_ent;
1020         res->dl_dx_leaf_bh = dx_leaf_bh;
1021         res->dl_dx_entry = dx_entry;
1022
1023         ret = 0;
1024 out:
1025         if (ret) {
1026                 brelse(dx_leaf_bh);
1027                 brelse(dir_ent_bh);
1028         }
1029         return ret;
1030 }
1031
1032 static int ocfs2_find_entry_dx(const char *name, int namelen,
1033                                struct inode *dir,
1034                                struct ocfs2_dir_lookup_result *lookup)
1035 {
1036         int ret;
1037         struct buffer_head *di_bh = NULL;
1038         struct ocfs2_dinode *di;
1039         struct buffer_head *dx_root_bh = NULL;
1040         struct ocfs2_dx_root_block *dx_root;
1041
1042         ret = ocfs2_read_inode_block(dir, &di_bh);
1043         if (ret) {
1044                 mlog_errno(ret);
1045                 goto out;
1046         }
1047
1048         di = (struct ocfs2_dinode *)di_bh->b_data;
1049
1050         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1051         if (ret) {
1052                 mlog_errno(ret);
1053                 goto out;
1054         }
1055         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1056
1057         ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1058         if (ret) {
1059                 if (ret != -ENOENT)
1060                         mlog_errno(ret);
1061                 goto out;
1062         }
1063
1064         lookup->dl_dx_root_bh = dx_root_bh;
1065         dx_root_bh = NULL;
1066 out:
1067         brelse(di_bh);
1068         brelse(dx_root_bh);
1069         return ret;
1070 }
1071
1072 /*
1073  * Try to find an entry of the provided name within 'dir'.
1074  *
1075  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1076  * returned and the struct 'res' will contain information useful to
1077  * other directory manipulation functions.
1078  *
1079  * Caller can NOT assume anything about the contents of the
1080  * buffer_heads - they are passed back only so that it can be passed
1081  * into any one of the manipulation functions (add entry, delete
1082  * entry, etc). As an example, bh in the extent directory case is a
1083  * data block, in the inline-data case it actually points to an inode,
1084  * in the indexed directory case, multiple buffers are involved.
1085  */
1086 int ocfs2_find_entry(const char *name, int namelen,
1087                      struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1088 {
1089         struct buffer_head *bh;
1090         struct ocfs2_dir_entry *res_dir = NULL;
1091
1092         if (ocfs2_dir_indexed(dir))
1093                 return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1094
1095         /*
1096          * The unindexed dir code only uses part of the lookup
1097          * structure, so there's no reason to push it down further
1098          * than this.
1099          */
1100         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1101                 bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1102         else
1103                 bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1104
1105         if (bh == NULL)
1106                 return -ENOENT;
1107
1108         lookup->dl_leaf_bh = bh;
1109         lookup->dl_entry = res_dir;
1110         return 0;
1111 }
1112
1113 /*
1114  * Update inode number and type of a previously found directory entry.
1115  */
1116 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1117                        struct ocfs2_dir_lookup_result *res,
1118                        struct inode *new_entry_inode)
1119 {
1120         int ret;
1121         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1122         struct ocfs2_dir_entry *de = res->dl_entry;
1123         struct buffer_head *de_bh = res->dl_leaf_bh;
1124
1125         /*
1126          * The same code works fine for both inline-data and extent
1127          * based directories, so no need to split this up.  The only
1128          * difference is the journal_access function.
1129          */
1130
1131         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1132                 access = ocfs2_journal_access_di;
1133
1134         ret = access(handle, INODE_CACHE(dir), de_bh,
1135                      OCFS2_JOURNAL_ACCESS_WRITE);
1136         if (ret) {
1137                 mlog_errno(ret);
1138                 goto out;
1139         }
1140
1141         de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1142         ocfs2_set_de_type(de, new_entry_inode->i_mode);
1143
1144         ocfs2_journal_dirty(handle, de_bh);
1145
1146 out:
1147         return ret;
1148 }
1149
1150 /*
1151  * __ocfs2_delete_entry deletes a directory entry by merging it with the
1152  * previous entry
1153  */
1154 static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1155                                 struct ocfs2_dir_entry *de_del,
1156                                 struct buffer_head *bh, char *first_de,
1157                                 unsigned int bytes)
1158 {
1159         struct ocfs2_dir_entry *de, *pde;
1160         int i, status = -ENOENT;
1161         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1162
1163         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1164                 access = ocfs2_journal_access_di;
1165
1166         i = 0;
1167         pde = NULL;
1168         de = (struct ocfs2_dir_entry *) first_de;
1169         while (i < bytes) {
1170                 if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1171                         status = -EIO;
1172                         mlog_errno(status);
1173                         goto bail;
1174                 }
1175                 if (de == de_del)  {
1176                         status = access(handle, INODE_CACHE(dir), bh,
1177                                         OCFS2_JOURNAL_ACCESS_WRITE);
1178                         if (status < 0) {
1179                                 status = -EIO;
1180                                 mlog_errno(status);
1181                                 goto bail;
1182                         }
1183                         if (pde)
1184                                 le16_add_cpu(&pde->rec_len,
1185                                                 le16_to_cpu(de->rec_len));
1186                         de->inode = 0;
1187                         dir->i_version++;
1188                         ocfs2_journal_dirty(handle, bh);
1189                         goto bail;
1190                 }
1191                 i += le16_to_cpu(de->rec_len);
1192                 pde = de;
1193                 de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1194         }
1195 bail:
1196         return status;
1197 }
1198
1199 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1200 {
1201         unsigned int hole;
1202
1203         if (le64_to_cpu(de->inode) == 0)
1204                 hole = le16_to_cpu(de->rec_len);
1205         else
1206                 hole = le16_to_cpu(de->rec_len) -
1207                         OCFS2_DIR_REC_LEN(de->name_len);
1208
1209         return hole;
1210 }
1211
1212 static int ocfs2_find_max_rec_len(struct super_block *sb,
1213                                   struct buffer_head *dirblock_bh)
1214 {
1215         int size, this_hole, largest_hole = 0;
1216         char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1217         struct ocfs2_dir_entry *de;
1218
1219         trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1220         size = ocfs2_dir_trailer_blk_off(sb);
1221         limit = start + size;
1222         de_buf = start;
1223         de = (struct ocfs2_dir_entry *)de_buf;
1224         do {
1225                 if (de_buf != trailer) {
1226                         this_hole = ocfs2_figure_dirent_hole(de);
1227                         if (this_hole > largest_hole)
1228                                 largest_hole = this_hole;
1229                 }
1230
1231                 de_buf += le16_to_cpu(de->rec_len);
1232                 de = (struct ocfs2_dir_entry *)de_buf;
1233         } while (de_buf < limit);
1234
1235         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1236                 return largest_hole;
1237         return 0;
1238 }
1239
1240 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1241                                        int index)
1242 {
1243         int num_used = le16_to_cpu(entry_list->de_num_used);
1244
1245         if (num_used == 1 || index == (num_used - 1))
1246                 goto clear;
1247
1248         memmove(&entry_list->de_entries[index],
1249                 &entry_list->de_entries[index + 1],
1250                 (num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1251 clear:
1252         num_used--;
1253         memset(&entry_list->de_entries[num_used], 0,
1254                sizeof(struct ocfs2_dx_entry));
1255         entry_list->de_num_used = cpu_to_le16(num_used);
1256 }
1257
1258 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1259                                  struct ocfs2_dir_lookup_result *lookup)
1260 {
1261         int ret, index, max_rec_len, add_to_free_list = 0;
1262         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1263         struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1264         struct ocfs2_dx_leaf *dx_leaf;
1265         struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1266         struct ocfs2_dir_block_trailer *trailer;
1267         struct ocfs2_dx_root_block *dx_root;
1268         struct ocfs2_dx_entry_list *entry_list;
1269
1270         /*
1271          * This function gets a bit messy because we might have to
1272          * modify the root block, regardless of whether the indexed
1273          * entries are stored inline.
1274          */
1275
1276         /*
1277          * *Only* set 'entry_list' here, based on where we're looking
1278          * for the indexed entries. Later, we might still want to
1279          * journal both blocks, based on free list state.
1280          */
1281         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1282         if (ocfs2_dx_root_inline(dx_root)) {
1283                 entry_list = &dx_root->dr_entries;
1284         } else {
1285                 dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1286                 entry_list = &dx_leaf->dl_list;
1287         }
1288
1289         /* Neither of these are a disk corruption - that should have
1290          * been caught by lookup, before we got here. */
1291         BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1292         BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1293
1294         index = (char *)dx_entry - (char *)entry_list->de_entries;
1295         index /= sizeof(*dx_entry);
1296
1297         if (index >= le16_to_cpu(entry_list->de_num_used)) {
1298                 mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1299                      (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1300                      entry_list, dx_entry);
1301                 return -EIO;
1302         }
1303
1304         /*
1305          * We know that removal of this dirent will leave enough room
1306          * for a new one, so add this block to the free list if it
1307          * isn't already there.
1308          */
1309         trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1310         if (trailer->db_free_rec_len == 0)
1311                 add_to_free_list = 1;
1312
1313         /*
1314          * Add the block holding our index into the journal before
1315          * removing the unindexed entry. If we get an error return
1316          * from __ocfs2_delete_entry(), then it hasn't removed the
1317          * entry yet. Likewise, successful return means we *must*
1318          * remove the indexed entry.
1319          *
1320          * We're also careful to journal the root tree block here as
1321          * the entry count needs to be updated. Also, we might be
1322          * adding to the start of the free list.
1323          */
1324         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1325                                       OCFS2_JOURNAL_ACCESS_WRITE);
1326         if (ret) {
1327                 mlog_errno(ret);
1328                 goto out;
1329         }
1330
1331         if (!ocfs2_dx_root_inline(dx_root)) {
1332                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1333                                               lookup->dl_dx_leaf_bh,
1334                                               OCFS2_JOURNAL_ACCESS_WRITE);
1335                 if (ret) {
1336                         mlog_errno(ret);
1337                         goto out;
1338                 }
1339         }
1340
1341         trace_ocfs2_delete_entry_dx((unsigned long long)OCFS2_I(dir)->ip_blkno,
1342                                     index);
1343
1344         ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1345                                    leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1346         if (ret) {
1347                 mlog_errno(ret);
1348                 goto out;
1349         }
1350
1351         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1352         trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1353         if (add_to_free_list) {
1354                 trailer->db_free_next = dx_root->dr_free_blk;
1355                 dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1356                 ocfs2_journal_dirty(handle, dx_root_bh);
1357         }
1358
1359         /* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1360         ocfs2_journal_dirty(handle, leaf_bh);
1361
1362         le32_add_cpu(&dx_root->dr_num_entries, -1);
1363         ocfs2_journal_dirty(handle, dx_root_bh);
1364
1365         ocfs2_dx_list_remove_entry(entry_list, index);
1366
1367         if (!ocfs2_dx_root_inline(dx_root))
1368                 ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1369
1370 out:
1371         return ret;
1372 }
1373
1374 static inline int ocfs2_delete_entry_id(handle_t *handle,
1375                                         struct inode *dir,
1376                                         struct ocfs2_dir_entry *de_del,
1377                                         struct buffer_head *bh)
1378 {
1379         int ret;
1380         struct buffer_head *di_bh = NULL;
1381         struct ocfs2_dinode *di;
1382         struct ocfs2_inline_data *data;
1383
1384         ret = ocfs2_read_inode_block(dir, &di_bh);
1385         if (ret) {
1386                 mlog_errno(ret);
1387                 goto out;
1388         }
1389
1390         di = (struct ocfs2_dinode *)di_bh->b_data;
1391         data = &di->id2.i_data;
1392
1393         ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1394                                    i_size_read(dir));
1395
1396         brelse(di_bh);
1397 out:
1398         return ret;
1399 }
1400
1401 static inline int ocfs2_delete_entry_el(handle_t *handle,
1402                                         struct inode *dir,
1403                                         struct ocfs2_dir_entry *de_del,
1404                                         struct buffer_head *bh)
1405 {
1406         return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1407                                     bh->b_size);
1408 }
1409
1410 /*
1411  * Delete a directory entry. Hide the details of directory
1412  * implementation from the caller.
1413  */
1414 int ocfs2_delete_entry(handle_t *handle,
1415                        struct inode *dir,
1416                        struct ocfs2_dir_lookup_result *res)
1417 {
1418         if (ocfs2_dir_indexed(dir))
1419                 return ocfs2_delete_entry_dx(handle, dir, res);
1420
1421         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1422                 return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1423                                              res->dl_leaf_bh);
1424
1425         return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1426                                      res->dl_leaf_bh);
1427 }
1428
1429 /*
1430  * Check whether 'de' has enough room to hold an entry of
1431  * 'new_rec_len' bytes.
1432  */
1433 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1434                                          unsigned int new_rec_len)
1435 {
1436         unsigned int de_really_used;
1437
1438         /* Check whether this is an empty record with enough space */
1439         if (le64_to_cpu(de->inode) == 0 &&
1440             le16_to_cpu(de->rec_len) >= new_rec_len)
1441                 return 1;
1442
1443         /*
1444          * Record might have free space at the end which we can
1445          * use.
1446          */
1447         de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1448         if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1449             return 1;
1450
1451         return 0;
1452 }
1453
1454 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1455                                           struct ocfs2_dx_entry *dx_new_entry)
1456 {
1457         int i;
1458
1459         i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1460         dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1461
1462         le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1463 }
1464
1465 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1466                                        struct ocfs2_dx_hinfo *hinfo,
1467                                        u64 dirent_blk)
1468 {
1469         int i;
1470         struct ocfs2_dx_entry *dx_entry;
1471
1472         i = le16_to_cpu(entry_list->de_num_used);
1473         dx_entry = &entry_list->de_entries[i];
1474
1475         memset(dx_entry, 0, sizeof(*dx_entry));
1476         dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1477         dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1478         dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1479
1480         le16_add_cpu(&entry_list->de_num_used, 1);
1481 }
1482
1483 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1484                                       struct ocfs2_dx_hinfo *hinfo,
1485                                       u64 dirent_blk,
1486                                       struct buffer_head *dx_leaf_bh)
1487 {
1488         int ret;
1489         struct ocfs2_dx_leaf *dx_leaf;
1490
1491         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1492                                       OCFS2_JOURNAL_ACCESS_WRITE);
1493         if (ret) {
1494                 mlog_errno(ret);
1495                 goto out;
1496         }
1497
1498         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1499         ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1500         ocfs2_journal_dirty(handle, dx_leaf_bh);
1501
1502 out:
1503         return ret;
1504 }
1505
1506 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1507                                         struct ocfs2_dx_hinfo *hinfo,
1508                                         u64 dirent_blk,
1509                                         struct ocfs2_dx_root_block *dx_root)
1510 {
1511         ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1512 }
1513
1514 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1515                                struct ocfs2_dir_lookup_result *lookup)
1516 {
1517         int ret = 0;
1518         struct ocfs2_dx_root_block *dx_root;
1519         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1520
1521         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1522                                       OCFS2_JOURNAL_ACCESS_WRITE);
1523         if (ret) {
1524                 mlog_errno(ret);
1525                 goto out;
1526         }
1527
1528         dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1529         if (ocfs2_dx_root_inline(dx_root)) {
1530                 ocfs2_dx_inline_root_insert(dir, handle,
1531                                             &lookup->dl_hinfo,
1532                                             lookup->dl_leaf_bh->b_blocknr,
1533                                             dx_root);
1534         } else {
1535                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1536                                                  lookup->dl_leaf_bh->b_blocknr,
1537                                                  lookup->dl_dx_leaf_bh);
1538                 if (ret)
1539                         goto out;
1540         }
1541
1542         le32_add_cpu(&dx_root->dr_num_entries, 1);
1543         ocfs2_journal_dirty(handle, dx_root_bh);
1544
1545 out:
1546         return ret;
1547 }
1548
1549 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1550                                        handle_t *handle,
1551                                        struct ocfs2_dir_lookup_result *lookup)
1552 {
1553         struct ocfs2_dir_block_trailer *trailer, *prev;
1554         struct ocfs2_dx_root_block *dx_root;
1555         struct buffer_head *bh;
1556
1557         trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1558
1559         if (ocfs2_free_list_at_root(lookup)) {
1560                 bh = lookup->dl_dx_root_bh;
1561                 dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1562                 dx_root->dr_free_blk = trailer->db_free_next;
1563         } else {
1564                 bh = lookup->dl_prev_leaf_bh;
1565                 prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1566                 prev->db_free_next = trailer->db_free_next;
1567         }
1568
1569         trailer->db_free_rec_len = cpu_to_le16(0);
1570         trailer->db_free_next = cpu_to_le64(0);
1571
1572         ocfs2_journal_dirty(handle, bh);
1573         ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1574 }
1575
1576 /*
1577  * This expects that a journal write has been reserved on
1578  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1579  */
1580 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1581                                    struct ocfs2_dir_lookup_result *lookup)
1582 {
1583         int max_rec_len;
1584         struct ocfs2_dir_block_trailer *trailer;
1585
1586         /* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1587         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1588         if (max_rec_len) {
1589                 /*
1590                  * There's still room in this block, so no need to remove it
1591                  * from the free list. In this case, we just want to update
1592                  * the rec len accounting.
1593                  */
1594                 trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1595                 trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1596                 ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1597         } else {
1598                 ocfs2_remove_block_from_free_list(dir, handle, lookup);
1599         }
1600 }
1601
1602 /* we don't always have a dentry for what we want to add, so people
1603  * like orphan dir can call this instead.
1604  *
1605  * The lookup context must have been filled from
1606  * ocfs2_prepare_dir_for_insert.
1607  */
1608 int __ocfs2_add_entry(handle_t *handle,
1609                       struct inode *dir,
1610                       const char *name, int namelen,
1611                       struct inode *inode, u64 blkno,
1612                       struct buffer_head *parent_fe_bh,
1613                       struct ocfs2_dir_lookup_result *lookup)
1614 {
1615         unsigned long offset;
1616         unsigned short rec_len;
1617         struct ocfs2_dir_entry *de, *de1;
1618         struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1619         struct super_block *sb = dir->i_sb;
1620         int retval, status;
1621         unsigned int size = sb->s_blocksize;
1622         struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1623         char *data_start = insert_bh->b_data;
1624
1625         if (!namelen)
1626                 return -EINVAL;
1627
1628         if (ocfs2_dir_indexed(dir)) {
1629                 struct buffer_head *bh;
1630
1631                 /*
1632                  * An indexed dir may require that we update the free space
1633                  * list. Reserve a write to the previous node in the list so
1634                  * that we don't fail later.
1635                  *
1636                  * XXX: This can be either a dx_root_block, or an unindexed
1637                  * directory tree leaf block.
1638                  */
1639                 if (ocfs2_free_list_at_root(lookup)) {
1640                         bh = lookup->dl_dx_root_bh;
1641                         retval = ocfs2_journal_access_dr(handle,
1642                                                  INODE_CACHE(dir), bh,
1643                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1644                 } else {
1645                         bh = lookup->dl_prev_leaf_bh;
1646                         retval = ocfs2_journal_access_db(handle,
1647                                                  INODE_CACHE(dir), bh,
1648                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1649                 }
1650                 if (retval) {
1651                         mlog_errno(retval);
1652                         return retval;
1653                 }
1654         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1655                 data_start = di->id2.i_data.id_data;
1656                 size = i_size_read(dir);
1657
1658                 BUG_ON(insert_bh != parent_fe_bh);
1659         }
1660
1661         rec_len = OCFS2_DIR_REC_LEN(namelen);
1662         offset = 0;
1663         de = (struct ocfs2_dir_entry *) data_start;
1664         while (1) {
1665                 BUG_ON((char *)de >= (size + data_start));
1666
1667                 /* These checks should've already been passed by the
1668                  * prepare function, but I guess we can leave them
1669                  * here anyway. */
1670                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1671                         retval = -ENOENT;
1672                         goto bail;
1673                 }
1674                 if (ocfs2_match(namelen, name, de)) {
1675                         retval = -EEXIST;
1676                         goto bail;
1677                 }
1678
1679                 /* We're guaranteed that we should have space, so we
1680                  * can't possibly have hit the trailer...right? */
1681                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1682                                 "Hit dir trailer trying to insert %.*s "
1683                                 "(namelen %d) into directory %llu.  "
1684                                 "offset is %lu, trailer offset is %d\n",
1685                                 namelen, name, namelen,
1686                                 (unsigned long long)parent_fe_bh->b_blocknr,
1687                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1688
1689                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1690                         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
1691                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1692                         if (retval < 0) {
1693                                 mlog_errno(retval);
1694                                 goto bail;
1695                         }
1696
1697                         if (insert_bh == parent_fe_bh)
1698                                 status = ocfs2_journal_access_di(handle,
1699                                                                  INODE_CACHE(dir),
1700                                                                  insert_bh,
1701                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1702                         else {
1703                                 status = ocfs2_journal_access_db(handle,
1704                                                                  INODE_CACHE(dir),
1705                                                                  insert_bh,
1706                                               OCFS2_JOURNAL_ACCESS_WRITE);
1707
1708                                 if (ocfs2_dir_indexed(dir)) {
1709                                         status = ocfs2_dx_dir_insert(dir,
1710                                                                 handle,
1711                                                                 lookup);
1712                                         if (status) {
1713                                                 mlog_errno(status);
1714                                                 goto bail;
1715                                         }
1716                                 }
1717                         }
1718
1719                         /* By now the buffer is marked for journaling */
1720                         offset += le16_to_cpu(de->rec_len);
1721                         if (le64_to_cpu(de->inode)) {
1722                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1723                                         OCFS2_DIR_REC_LEN(de->name_len));
1724                                 de1->rec_len =
1725                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1726                                         OCFS2_DIR_REC_LEN(de->name_len));
1727                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1728                                 de = de1;
1729                         }
1730                         de->file_type = OCFS2_FT_UNKNOWN;
1731                         if (blkno) {
1732                                 de->inode = cpu_to_le64(blkno);
1733                                 ocfs2_set_de_type(de, inode->i_mode);
1734                         } else
1735                                 de->inode = 0;
1736                         de->name_len = namelen;
1737                         memcpy(de->name, name, namelen);
1738
1739                         if (ocfs2_dir_indexed(dir))
1740                                 ocfs2_recalc_free_list(dir, handle, lookup);
1741
1742                         dir->i_version++;
1743                         ocfs2_journal_dirty(handle, insert_bh);
1744                         retval = 0;
1745                         goto bail;
1746                 }
1747
1748                 offset += le16_to_cpu(de->rec_len);
1749                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1750         }
1751
1752         /* when you think about it, the assert above should prevent us
1753          * from ever getting here. */
1754         retval = -ENOSPC;
1755 bail:
1756         if (retval)
1757                 mlog_errno(retval);
1758
1759         return retval;
1760 }
1761
1762 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1763                                     u64 *f_version,
1764                                     struct dir_context *ctx)
1765 {
1766         int ret, i;
1767         unsigned long offset = ctx->pos;
1768         struct buffer_head *di_bh = NULL;
1769         struct ocfs2_dinode *di;
1770         struct ocfs2_inline_data *data;
1771         struct ocfs2_dir_entry *de;
1772
1773         ret = ocfs2_read_inode_block(inode, &di_bh);
1774         if (ret) {
1775                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1776                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1777                 goto out;
1778         }
1779
1780         di = (struct ocfs2_dinode *)di_bh->b_data;
1781         data = &di->id2.i_data;
1782
1783         while (ctx->pos < i_size_read(inode)) {
1784                 /* If the dir block has changed since the last call to
1785                  * readdir(2), then we might be pointing to an invalid
1786                  * dirent right now.  Scan from the start of the block
1787                  * to make sure. */
1788                 if (*f_version != inode->i_version) {
1789                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1790                                 de = (struct ocfs2_dir_entry *)
1791                                         (data->id_data + i);
1792                                 /* It's too expensive to do a full
1793                                  * dirent test each time round this
1794                                  * loop, but we do have to test at
1795                                  * least that it is non-zero.  A
1796                                  * failure will be detected in the
1797                                  * dirent test below. */
1798                                 if (le16_to_cpu(de->rec_len) <
1799                                     OCFS2_DIR_REC_LEN(1))
1800                                         break;
1801                                 i += le16_to_cpu(de->rec_len);
1802                         }
1803                         ctx->pos = offset = i;
1804                         *f_version = inode->i_version;
1805                 }
1806
1807                 de = (struct ocfs2_dir_entry *) (data->id_data + ctx->pos);
1808                 if (!ocfs2_check_dir_entry(inode, de, di_bh, ctx->pos)) {
1809                         /* On error, skip the f_pos to the end. */
1810                         ctx->pos = i_size_read(inode);
1811                         break;
1812                 }
1813                 offset += le16_to_cpu(de->rec_len);
1814                 if (le64_to_cpu(de->inode)) {
1815                         unsigned char d_type = DT_UNKNOWN;
1816
1817                         if (de->file_type < OCFS2_FT_MAX)
1818                                 d_type = ocfs2_filetype_table[de->file_type];
1819
1820                         if (!dir_emit(ctx, de->name, de->name_len,
1821                                       le64_to_cpu(de->inode), d_type))
1822                                 goto out;
1823                 }
1824                 ctx->pos += le16_to_cpu(de->rec_len);
1825         }
1826 out:
1827         brelse(di_bh);
1828         return 0;
1829 }
1830
1831 /*
1832  * NOTE: This function can be called against unindexed directories,
1833  * and indexed ones.
1834  */
1835 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1836                                     u64 *f_version,
1837                                     struct dir_context *ctx,
1838                                     bool persist)
1839 {
1840         unsigned long offset, blk, last_ra_blk = 0;
1841         int i;
1842         struct buffer_head * bh, * tmp;
1843         struct ocfs2_dir_entry * de;
1844         struct super_block * sb = inode->i_sb;
1845         unsigned int ra_sectors = 16;
1846         int stored = 0;
1847
1848         bh = NULL;
1849
1850         offset = ctx->pos & (sb->s_blocksize - 1);
1851
1852         while (ctx->pos < i_size_read(inode)) {
1853                 blk = ctx->pos >> sb->s_blocksize_bits;
1854                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1855                         /* Skip the corrupt dirblock and keep trying */
1856                         ctx->pos += sb->s_blocksize - offset;
1857                         continue;
1858                 }
1859
1860                 /* The idea here is to begin with 8k read-ahead and to stay
1861                  * 4k ahead of our current position.
1862                  *
1863                  * TODO: Use the pagecache for this. We just need to
1864                  * make sure it's cluster-safe... */
1865                 if (!last_ra_blk
1866                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1867                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1868                              i > 0; i--) {
1869                                 tmp = NULL;
1870                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1871                                                           OCFS2_BH_READAHEAD))
1872                                         brelse(tmp);
1873                         }
1874                         last_ra_blk = blk;
1875                         ra_sectors = 8;
1876                 }
1877
1878                 /* If the dir block has changed since the last call to
1879                  * readdir(2), then we might be pointing to an invalid
1880                  * dirent right now.  Scan from the start of the block
1881                  * to make sure. */
1882                 if (*f_version != inode->i_version) {
1883                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1884                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1885                                 /* It's too expensive to do a full
1886                                  * dirent test each time round this
1887                                  * loop, but we do have to test at
1888                                  * least that it is non-zero.  A
1889                                  * failure will be detected in the
1890                                  * dirent test below. */
1891                                 if (le16_to_cpu(de->rec_len) <
1892                                     OCFS2_DIR_REC_LEN(1))
1893                                         break;
1894                                 i += le16_to_cpu(de->rec_len);
1895                         }
1896                         offset = i;
1897                         ctx->pos = (ctx->pos & ~(sb->s_blocksize - 1))
1898                                 | offset;
1899                         *f_version = inode->i_version;
1900                 }
1901
1902                 while (ctx->pos < i_size_read(inode)
1903                        && offset < sb->s_blocksize) {
1904                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1905                         if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1906                                 /* On error, skip the f_pos to the
1907                                    next block. */
1908                                 ctx->pos = (ctx->pos | (sb->s_blocksize - 1)) + 1;
1909                                 brelse(bh);
1910                                 continue;
1911                         }
1912                         if (le64_to_cpu(de->inode)) {
1913                                 unsigned char d_type = DT_UNKNOWN;
1914
1915                                 if (de->file_type < OCFS2_FT_MAX)
1916                                         d_type = ocfs2_filetype_table[de->file_type];
1917                                 if (!dir_emit(ctx, de->name,
1918                                                 de->name_len,
1919                                                 le64_to_cpu(de->inode),
1920                                                 d_type)) {
1921                                         brelse(bh);
1922                                         return 0;
1923                                 }
1924                                 stored++;
1925                         }
1926                         offset += le16_to_cpu(de->rec_len);
1927                         ctx->pos += le16_to_cpu(de->rec_len);
1928                 }
1929                 offset = 0;
1930                 brelse(bh);
1931                 bh = NULL;
1932                 if (!persist && stored)
1933                         break;
1934         }
1935         return 0;
1936 }
1937
1938 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1939                                  struct dir_context *ctx,
1940                                  bool persist)
1941 {
1942         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1943                 return ocfs2_dir_foreach_blk_id(inode, f_version, ctx);
1944         return ocfs2_dir_foreach_blk_el(inode, f_version, ctx, persist);
1945 }
1946
1947 /*
1948  * This is intended to be called from inside other kernel functions,
1949  * so we fake some arguments.
1950  */
1951 int ocfs2_dir_foreach(struct inode *inode, struct dir_context *ctx)
1952 {
1953         u64 version = inode->i_version;
1954         ocfs2_dir_foreach_blk(inode, &version, ctx, true);
1955         return 0;
1956 }
1957
1958 /*
1959  * ocfs2_readdir()
1960  *
1961  */
1962 int ocfs2_readdir(struct file *file, struct dir_context *ctx)
1963 {
1964         int error = 0;
1965         struct inode *inode = file_inode(file);
1966         int lock_level = 0;
1967
1968         trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
1969
1970         error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level);
1971         if (lock_level && error >= 0) {
1972                 /* We release EX lock which used to update atime
1973                  * and get PR lock again to reduce contention
1974                  * on commonly accessed directories. */
1975                 ocfs2_inode_unlock(inode, 1);
1976                 lock_level = 0;
1977                 error = ocfs2_inode_lock(inode, NULL, 0);
1978         }
1979         if (error < 0) {
1980                 if (error != -ENOENT)
1981                         mlog_errno(error);
1982                 /* we haven't got any yet, so propagate the error. */
1983                 goto bail_nolock;
1984         }
1985
1986         error = ocfs2_dir_foreach_blk(inode, &file->f_version, ctx, false);
1987
1988         ocfs2_inode_unlock(inode, lock_level);
1989         if (error)
1990                 mlog_errno(error);
1991
1992 bail_nolock:
1993
1994         return error;
1995 }
1996
1997 /*
1998  * NOTE: this should always be called with parent dir i_mutex taken.
1999  */
2000 int ocfs2_find_files_on_disk(const char *name,
2001                              int namelen,
2002                              u64 *blkno,
2003                              struct inode *inode,
2004                              struct ocfs2_dir_lookup_result *lookup)
2005 {
2006         int status = -ENOENT;
2007
2008         trace_ocfs2_find_files_on_disk(namelen, name, blkno,
2009                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
2010
2011         status = ocfs2_find_entry(name, namelen, inode, lookup);
2012         if (status)
2013                 goto leave;
2014
2015         *blkno = le64_to_cpu(lookup->dl_entry->inode);
2016
2017         status = 0;
2018 leave:
2019
2020         return status;
2021 }
2022
2023 /*
2024  * Convenience function for callers which just want the block number
2025  * mapped to a name and don't require the full dirent info, etc.
2026  */
2027 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
2028                                int namelen, u64 *blkno)
2029 {
2030         int ret;
2031         struct ocfs2_dir_lookup_result lookup = { NULL, };
2032
2033         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2034         ocfs2_free_dir_lookup_result(&lookup);
2035
2036         return ret;
2037 }
2038
2039 /* Check for a name within a directory.
2040  *
2041  * Return 0 if the name does not exist
2042  * Return -EEXIST if the directory contains the name
2043  *
2044  * Callers should have i_mutex + a cluster lock on dir
2045  */
2046 int ocfs2_check_dir_for_entry(struct inode *dir,
2047                               const char *name,
2048                               int namelen)
2049 {
2050         int ret;
2051         struct ocfs2_dir_lookup_result lookup = { NULL, };
2052
2053         trace_ocfs2_check_dir_for_entry(
2054                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2055
2056         ret = -EEXIST;
2057         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0)
2058                 goto bail;
2059
2060         ret = 0;
2061 bail:
2062         ocfs2_free_dir_lookup_result(&lookup);
2063
2064         if (ret)
2065                 mlog_errno(ret);
2066         return ret;
2067 }
2068
2069 struct ocfs2_empty_dir_priv {
2070         struct dir_context ctx;
2071         unsigned seen_dot;
2072         unsigned seen_dot_dot;
2073         unsigned seen_other;
2074         unsigned dx_dir;
2075 };
2076 static int ocfs2_empty_dir_filldir(struct dir_context *ctx, const char *name,
2077                                    int name_len, loff_t pos, u64 ino,
2078                                    unsigned type)
2079 {
2080         struct ocfs2_empty_dir_priv *p =
2081                 container_of(ctx, struct ocfs2_empty_dir_priv, ctx);
2082
2083         /*
2084          * Check the positions of "." and ".." records to be sure
2085          * they're in the correct place.
2086          *
2087          * Indexed directories don't need to proceed past the first
2088          * two entries, so we end the scan after seeing '..'. Despite
2089          * that, we allow the scan to proceed In the event that we
2090          * have a corrupted indexed directory (no dot or dot dot
2091          * entries). This allows us to double check for existing
2092          * entries which might not have been found in the index.
2093          */
2094         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2095                 p->seen_dot = 1;
2096                 return 0;
2097         }
2098
2099         if (name_len == 2 && !strncmp("..", name, 2) &&
2100             pos == OCFS2_DIR_REC_LEN(1)) {
2101                 p->seen_dot_dot = 1;
2102
2103                 if (p->dx_dir && p->seen_dot)
2104                         return 1;
2105
2106                 return 0;
2107         }
2108
2109         p->seen_other = 1;
2110         return 1;
2111 }
2112
2113 static int ocfs2_empty_dir_dx(struct inode *inode,
2114                               struct ocfs2_empty_dir_priv *priv)
2115 {
2116         int ret;
2117         struct buffer_head *di_bh = NULL;
2118         struct buffer_head *dx_root_bh = NULL;
2119         struct ocfs2_dinode *di;
2120         struct ocfs2_dx_root_block *dx_root;
2121
2122         priv->dx_dir = 1;
2123
2124         ret = ocfs2_read_inode_block(inode, &di_bh);
2125         if (ret) {
2126                 mlog_errno(ret);
2127                 goto out;
2128         }
2129         di = (struct ocfs2_dinode *)di_bh->b_data;
2130
2131         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2132         if (ret) {
2133                 mlog_errno(ret);
2134                 goto out;
2135         }
2136         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2137
2138         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2139                 priv->seen_other = 1;
2140
2141 out:
2142         brelse(di_bh);
2143         brelse(dx_root_bh);
2144         return ret;
2145 }
2146
2147 /*
2148  * routine to check that the specified directory is empty (for rmdir)
2149  *
2150  * Returns 1 if dir is empty, zero otherwise.
2151  *
2152  * XXX: This is a performance problem for unindexed directories.
2153  */
2154 int ocfs2_empty_dir(struct inode *inode)
2155 {
2156         int ret;
2157         struct ocfs2_empty_dir_priv priv = {
2158                 .ctx.actor = ocfs2_empty_dir_filldir,
2159         };
2160
2161         if (ocfs2_dir_indexed(inode)) {
2162                 ret = ocfs2_empty_dir_dx(inode, &priv);
2163                 if (ret)
2164                         mlog_errno(ret);
2165                 /*
2166                  * We still run ocfs2_dir_foreach to get the checks
2167                  * for "." and "..".
2168                  */
2169         }
2170
2171         ret = ocfs2_dir_foreach(inode, &priv.ctx);
2172         if (ret)
2173                 mlog_errno(ret);
2174
2175         if (!priv.seen_dot || !priv.seen_dot_dot) {
2176                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2177                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2178                 /*
2179                  * XXX: Is it really safe to allow an unlink to continue?
2180                  */
2181                 return 1;
2182         }
2183
2184         return !priv.seen_other;
2185 }
2186
2187 /*
2188  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2189  * "..", which might be used during creation of a directory with a trailing
2190  * header. It is otherwise safe to ignore the return code.
2191  */
2192 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2193                                                           struct inode *parent,
2194                                                           char *start,
2195                                                           unsigned int size)
2196 {
2197         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2198
2199         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2200         de->name_len = 1;
2201         de->rec_len =
2202                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2203         strcpy(de->name, ".");
2204         ocfs2_set_de_type(de, S_IFDIR);
2205
2206         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2207         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2208         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2209         de->name_len = 2;
2210         strcpy(de->name, "..");
2211         ocfs2_set_de_type(de, S_IFDIR);
2212
2213         return de;
2214 }
2215
2216 /*
2217  * This works together with code in ocfs2_mknod_locked() which sets
2218  * the inline-data flag and initializes the inline-data section.
2219  */
2220 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2221                                  handle_t *handle,
2222                                  struct inode *parent,
2223                                  struct inode *inode,
2224                                  struct buffer_head *di_bh)
2225 {
2226         int ret;
2227         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2228         struct ocfs2_inline_data *data = &di->id2.i_data;
2229         unsigned int size = le16_to_cpu(data->id_count);
2230
2231         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2232                                       OCFS2_JOURNAL_ACCESS_WRITE);
2233         if (ret) {
2234                 mlog_errno(ret);
2235                 goto out;
2236         }
2237
2238         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2239         ocfs2_journal_dirty(handle, di_bh);
2240
2241         i_size_write(inode, size);
2242         set_nlink(inode, 2);
2243         inode->i_blocks = ocfs2_inode_sector_count(inode);
2244
2245         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2246         if (ret < 0)
2247                 mlog_errno(ret);
2248
2249 out:
2250         return ret;
2251 }
2252
2253 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2254                                  handle_t *handle,
2255                                  struct inode *parent,
2256                                  struct inode *inode,
2257                                  struct buffer_head *fe_bh,
2258                                  struct ocfs2_alloc_context *data_ac,
2259                                  struct buffer_head **ret_new_bh)
2260 {
2261         int status;
2262         unsigned int size = osb->sb->s_blocksize;
2263         struct buffer_head *new_bh = NULL;
2264         struct ocfs2_dir_entry *de;
2265
2266         if (ocfs2_new_dir_wants_trailer(inode))
2267                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2268
2269         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2270                                      data_ac, NULL, &new_bh);
2271         if (status < 0) {
2272                 mlog_errno(status);
2273                 goto bail;
2274         }
2275
2276         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2277
2278         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2279                                          OCFS2_JOURNAL_ACCESS_CREATE);
2280         if (status < 0) {
2281                 mlog_errno(status);
2282                 goto bail;
2283         }
2284         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2285
2286         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2287         if (ocfs2_new_dir_wants_trailer(inode)) {
2288                 int size = le16_to_cpu(de->rec_len);
2289
2290                 /*
2291                  * Figure out the size of the hole left over after
2292                  * insertion of '.' and '..'. The trailer wants this
2293                  * information.
2294                  */
2295                 size -= OCFS2_DIR_REC_LEN(2);
2296                 size -= sizeof(struct ocfs2_dir_block_trailer);
2297
2298                 ocfs2_init_dir_trailer(inode, new_bh, size);
2299         }
2300
2301         ocfs2_journal_dirty(handle, new_bh);
2302
2303         i_size_write(inode, inode->i_sb->s_blocksize);
2304         set_nlink(inode, 2);
2305         inode->i_blocks = ocfs2_inode_sector_count(inode);
2306         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2307         if (status < 0) {
2308                 mlog_errno(status);
2309                 goto bail;
2310         }
2311
2312         status = 0;
2313         if (ret_new_bh) {
2314                 *ret_new_bh = new_bh;
2315                 new_bh = NULL;
2316         }
2317 bail:
2318         brelse(new_bh);
2319
2320         return status;
2321 }
2322
2323 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2324                                      handle_t *handle, struct inode *dir,
2325                                      struct buffer_head *di_bh,
2326                                      struct buffer_head *dirdata_bh,
2327                                      struct ocfs2_alloc_context *meta_ac,
2328                                      int dx_inline, u32 num_entries,
2329                                      struct buffer_head **ret_dx_root_bh)
2330 {
2331         int ret;
2332         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2333         u16 dr_suballoc_bit;
2334         u64 suballoc_loc, dr_blkno;
2335         unsigned int num_bits;
2336         struct buffer_head *dx_root_bh = NULL;
2337         struct ocfs2_dx_root_block *dx_root;
2338         struct ocfs2_dir_block_trailer *trailer =
2339                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2340
2341         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2342                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2343         if (ret) {
2344                 mlog_errno(ret);
2345                 goto out;
2346         }
2347
2348         trace_ocfs2_dx_dir_attach_index(
2349                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2350                                 (unsigned long long)dr_blkno);
2351
2352         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2353         if (dx_root_bh == NULL) {
2354                 ret = -ENOMEM;
2355                 goto out;
2356         }
2357         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2358
2359         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2360                                       OCFS2_JOURNAL_ACCESS_CREATE);
2361         if (ret < 0) {
2362                 mlog_errno(ret);
2363                 goto out;
2364         }
2365
2366         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2367         memset(dx_root, 0, osb->sb->s_blocksize);
2368         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2369         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2370         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2371         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2372         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2373         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2374         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2375         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2376         if (le16_to_cpu(trailer->db_free_rec_len))
2377                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2378         else
2379                 dx_root->dr_free_blk = cpu_to_le64(0);
2380
2381         if (dx_inline) {
2382                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2383                 dx_root->dr_entries.de_count =
2384                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2385         } else {
2386                 dx_root->dr_list.l_count =
2387                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2388         }
2389         ocfs2_journal_dirty(handle, dx_root_bh);
2390
2391         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2392                                       OCFS2_JOURNAL_ACCESS_CREATE);
2393         if (ret) {
2394                 mlog_errno(ret);
2395                 goto out;
2396         }
2397
2398         di->i_dx_root = cpu_to_le64(dr_blkno);
2399
2400         spin_lock(&OCFS2_I(dir)->ip_lock);
2401         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2402         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2403         spin_unlock(&OCFS2_I(dir)->ip_lock);
2404
2405         ocfs2_journal_dirty(handle, di_bh);
2406
2407         *ret_dx_root_bh = dx_root_bh;
2408         dx_root_bh = NULL;
2409
2410 out:
2411         brelse(dx_root_bh);
2412         return ret;
2413 }
2414
2415 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2416                                        handle_t *handle, struct inode *dir,
2417                                        struct buffer_head **dx_leaves,
2418                                        int num_dx_leaves, u64 start_blk)
2419 {
2420         int ret, i;
2421         struct ocfs2_dx_leaf *dx_leaf;
2422         struct buffer_head *bh;
2423
2424         for (i = 0; i < num_dx_leaves; i++) {
2425                 bh = sb_getblk(osb->sb, start_blk + i);
2426                 if (bh == NULL) {
2427                         ret = -ENOMEM;
2428                         goto out;
2429                 }
2430                 dx_leaves[i] = bh;
2431
2432                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2433
2434                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2435                                               OCFS2_JOURNAL_ACCESS_CREATE);
2436                 if (ret < 0) {
2437                         mlog_errno(ret);
2438                         goto out;
2439                 }
2440
2441                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2442
2443                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2444                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2445                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2446                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2447                 dx_leaf->dl_list.de_count =
2448                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2449
2450                 trace_ocfs2_dx_dir_format_cluster(
2451                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2452                                 (unsigned long long)bh->b_blocknr,
2453                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2454
2455                 ocfs2_journal_dirty(handle, bh);
2456         }
2457
2458         ret = 0;
2459 out:
2460         return ret;
2461 }
2462
2463 /*
2464  * Allocates and formats a new cluster for use in an indexed dir
2465  * leaf. This version will not do the extent insert, so that it can be
2466  * used by operations which need careful ordering.
2467  */
2468 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2469                                       u32 cpos, handle_t *handle,
2470                                       struct ocfs2_alloc_context *data_ac,
2471                                       struct buffer_head **dx_leaves,
2472                                       int num_dx_leaves, u64 *ret_phys_blkno)
2473 {
2474         int ret;
2475         u32 phys, num;
2476         u64 phys_blkno;
2477         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2478
2479         /*
2480          * XXX: For create, this should claim cluster for the index
2481          * *before* the unindexed insert so that we have a better
2482          * chance of contiguousness as the directory grows in number
2483          * of entries.
2484          */
2485         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2486         if (ret) {
2487                 mlog_errno(ret);
2488                 goto out;
2489         }
2490
2491         /*
2492          * Format the new cluster first. That way, we're inserting
2493          * valid data.
2494          */
2495         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2496         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2497                                           num_dx_leaves, phys_blkno);
2498         if (ret) {
2499                 mlog_errno(ret);
2500                 goto out;
2501         }
2502
2503         *ret_phys_blkno = phys_blkno;
2504 out:
2505         return ret;
2506 }
2507
2508 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2509                                     struct ocfs2_extent_tree *et,
2510                                     u32 cpos, handle_t *handle,
2511                                     struct ocfs2_alloc_context *data_ac,
2512                                     struct ocfs2_alloc_context *meta_ac,
2513                                     struct buffer_head **dx_leaves,
2514                                     int num_dx_leaves)
2515 {
2516         int ret;
2517         u64 phys_blkno;
2518
2519         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2520                                          num_dx_leaves, &phys_blkno);
2521         if (ret) {
2522                 mlog_errno(ret);
2523                 goto out;
2524         }
2525
2526         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2527                                   meta_ac);
2528         if (ret)
2529                 mlog_errno(ret);
2530 out:
2531         return ret;
2532 }
2533
2534 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2535                                                         int *ret_num_leaves)
2536 {
2537         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2538         struct buffer_head **dx_leaves;
2539
2540         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2541                             GFP_NOFS);
2542         if (dx_leaves && ret_num_leaves)
2543                 *ret_num_leaves = num_dx_leaves;
2544
2545         return dx_leaves;
2546 }
2547
2548 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2549                                  handle_t *handle,
2550                                  struct inode *parent,
2551                                  struct inode *inode,
2552                                  struct buffer_head *di_bh,
2553                                  struct ocfs2_alloc_context *data_ac,
2554                                  struct ocfs2_alloc_context *meta_ac)
2555 {
2556         int ret;
2557         struct buffer_head *leaf_bh = NULL;
2558         struct buffer_head *dx_root_bh = NULL;
2559         struct ocfs2_dx_hinfo hinfo;
2560         struct ocfs2_dx_root_block *dx_root;
2561         struct ocfs2_dx_entry_list *entry_list;
2562
2563         /*
2564          * Our strategy is to create the directory as though it were
2565          * unindexed, then add the index block. This works with very
2566          * little complication since the state of a new directory is a
2567          * very well known quantity.
2568          *
2569          * Essentially, we have two dirents ("." and ".."), in the 1st
2570          * block which need indexing. These are easily inserted into
2571          * the index block.
2572          */
2573
2574         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2575                                     data_ac, &leaf_bh);
2576         if (ret) {
2577                 mlog_errno(ret);
2578                 goto out;
2579         }
2580
2581         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2582                                         meta_ac, 1, 2, &dx_root_bh);
2583         if (ret) {
2584                 mlog_errno(ret);
2585                 goto out;
2586         }
2587         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2588         entry_list = &dx_root->dr_entries;
2589
2590         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2591         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2592         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2593
2594         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2595         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2596
2597 out:
2598         brelse(dx_root_bh);
2599         brelse(leaf_bh);
2600         return ret;
2601 }
2602
2603 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2604                        handle_t *handle,
2605                        struct inode *parent,
2606                        struct inode *inode,
2607                        struct buffer_head *fe_bh,
2608                        struct ocfs2_alloc_context *data_ac,
2609                        struct ocfs2_alloc_context *meta_ac)
2610
2611 {
2612         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2613
2614         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2615                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2616
2617         if (ocfs2_supports_indexed_dirs(osb))
2618                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2619                                              data_ac, meta_ac);
2620
2621         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2622                                      data_ac, NULL);
2623 }
2624
2625 static int ocfs2_dx_dir_index_block(struct inode *dir,
2626                                     handle_t *handle,
2627                                     struct buffer_head **dx_leaves,
2628                                     int num_dx_leaves,
2629                                     u32 *num_dx_entries,
2630                                     struct buffer_head *dirent_bh)
2631 {
2632         int ret = 0, namelen, i;
2633         char *de_buf, *limit;
2634         struct ocfs2_dir_entry *de;
2635         struct buffer_head *dx_leaf_bh;
2636         struct ocfs2_dx_hinfo hinfo;
2637         u64 dirent_blk = dirent_bh->b_blocknr;
2638
2639         de_buf = dirent_bh->b_data;
2640         limit = de_buf + dir->i_sb->s_blocksize;
2641
2642         while (de_buf < limit) {
2643                 de = (struct ocfs2_dir_entry *)de_buf;
2644
2645                 namelen = de->name_len;
2646                 if (!namelen || !de->inode)
2647                         goto inc;
2648
2649                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2650
2651                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2652                 dx_leaf_bh = dx_leaves[i];
2653
2654                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2655                                                  dirent_blk, dx_leaf_bh);
2656                 if (ret) {
2657                         mlog_errno(ret);
2658                         goto out;
2659                 }
2660
2661                 *num_dx_entries = *num_dx_entries + 1;
2662
2663 inc:
2664                 de_buf += le16_to_cpu(de->rec_len);
2665         }
2666
2667 out:
2668         return ret;
2669 }
2670
2671 /*
2672  * XXX: This expects dx_root_bh to already be part of the transaction.
2673  */
2674 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2675                                          struct buffer_head *dx_root_bh,
2676                                          struct buffer_head *dirent_bh)
2677 {
2678         char *de_buf, *limit;
2679         struct ocfs2_dx_root_block *dx_root;
2680         struct ocfs2_dir_entry *de;
2681         struct ocfs2_dx_hinfo hinfo;
2682         u64 dirent_blk = dirent_bh->b_blocknr;
2683
2684         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2685
2686         de_buf = dirent_bh->b_data;
2687         limit = de_buf + dir->i_sb->s_blocksize;
2688
2689         while (de_buf < limit) {
2690                 de = (struct ocfs2_dir_entry *)de_buf;
2691
2692                 if (!de->name_len || !de->inode)
2693                         goto inc;
2694
2695                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2696
2697                 trace_ocfs2_dx_dir_index_root_block(
2698                                 (unsigned long long)dir->i_ino,
2699                                 hinfo.major_hash, hinfo.minor_hash,
2700                                 de->name_len, de->name,
2701                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2702
2703                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2704                                            dirent_blk);
2705
2706                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2707 inc:
2708                 de_buf += le16_to_cpu(de->rec_len);
2709         }
2710 }
2711
2712 /*
2713  * Count the number of inline directory entries in di_bh and compare
2714  * them against the number of entries we can hold in an inline dx root
2715  * block.
2716  */
2717 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2718                                          struct buffer_head *di_bh)
2719 {
2720         int dirent_count = 0;
2721         char *de_buf, *limit;
2722         struct ocfs2_dir_entry *de;
2723         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2724
2725         de_buf = di->id2.i_data.id_data;
2726         limit = de_buf + i_size_read(dir);
2727
2728         while (de_buf < limit) {
2729                 de = (struct ocfs2_dir_entry *)de_buf;
2730
2731                 if (de->name_len && de->inode)
2732                         dirent_count++;
2733
2734                 de_buf += le16_to_cpu(de->rec_len);
2735         }
2736
2737         /* We are careful to leave room for one extra record. */
2738         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2739 }
2740
2741 /*
2742  * Expand rec_len of the rightmost dirent in a directory block so that it
2743  * contains the end of our valid space for dirents. We do this during
2744  * expansion from an inline directory to one with extents. The first dir block
2745  * in that case is taken from the inline data portion of the inode block.
2746  *
2747  * This will also return the largest amount of contiguous space for a dirent
2748  * in the block. That value is *not* necessarily the last dirent, even after
2749  * expansion. The directory indexing code wants this value for free space
2750  * accounting. We do this here since we're already walking the entire dir
2751  * block.
2752  *
2753  * We add the dir trailer if this filesystem wants it.
2754  */
2755 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2756                                              struct inode *dir)
2757 {
2758         struct super_block *sb = dir->i_sb;
2759         struct ocfs2_dir_entry *de;
2760         struct ocfs2_dir_entry *prev_de;
2761         char *de_buf, *limit;
2762         unsigned int new_size = sb->s_blocksize;
2763         unsigned int bytes, this_hole;
2764         unsigned int largest_hole = 0;
2765
2766         if (ocfs2_new_dir_wants_trailer(dir))
2767                 new_size = ocfs2_dir_trailer_blk_off(sb);
2768
2769         bytes = new_size - old_size;
2770
2771         limit = start + old_size;
2772         de_buf = start;
2773         de = (struct ocfs2_dir_entry *)de_buf;
2774         do {
2775                 this_hole = ocfs2_figure_dirent_hole(de);
2776                 if (this_hole > largest_hole)
2777                         largest_hole = this_hole;
2778
2779                 prev_de = de;
2780                 de_buf += le16_to_cpu(de->rec_len);
2781                 de = (struct ocfs2_dir_entry *)de_buf;
2782         } while (de_buf < limit);
2783
2784         le16_add_cpu(&prev_de->rec_len, bytes);
2785
2786         /* We need to double check this after modification of the final
2787          * dirent. */
2788         this_hole = ocfs2_figure_dirent_hole(prev_de);
2789         if (this_hole > largest_hole)
2790                 largest_hole = this_hole;
2791
2792         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2793                 return largest_hole;
2794         return 0;
2795 }
2796
2797 /*
2798  * We allocate enough clusters to fulfill "blocks_wanted", but set
2799  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2800  * rest automatically for us.
2801  *
2802  * *first_block_bh is a pointer to the 1st data block allocated to the
2803  *  directory.
2804  */
2805 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2806                                    unsigned int blocks_wanted,
2807                                    struct ocfs2_dir_lookup_result *lookup,
2808                                    struct buffer_head **first_block_bh)
2809 {
2810         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2811         struct super_block *sb = dir->i_sb;
2812         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2813                 credits = ocfs2_inline_to_extents_credits(sb);
2814         u64 dx_insert_blkno, blkno,
2815                 bytes = blocks_wanted << sb->s_blocksize_bits;
2816         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2817         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2818         struct ocfs2_alloc_context *data_ac = NULL;
2819         struct ocfs2_alloc_context *meta_ac = NULL;
2820         struct buffer_head *dirdata_bh = NULL;
2821         struct buffer_head *dx_root_bh = NULL;
2822         struct buffer_head **dx_leaves = NULL;
2823         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2824         handle_t *handle;
2825         struct ocfs2_extent_tree et;
2826         struct ocfs2_extent_tree dx_et;
2827         int did_quota = 0, bytes_allocated = 0;
2828
2829         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2830
2831         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2832         dx_alloc = 0;
2833
2834         down_write(&oi->ip_alloc_sem);
2835
2836         if (ocfs2_supports_indexed_dirs(osb)) {
2837                 credits += ocfs2_add_dir_index_credits(sb);
2838
2839                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2840                 if (!dx_inline) {
2841                         /* Add one more cluster for an index leaf */
2842                         dx_alloc++;
2843                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2844                                                                 &num_dx_leaves);
2845                         if (!dx_leaves) {
2846                                 ret = -ENOMEM;
2847                                 mlog_errno(ret);
2848                                 goto out;
2849                         }
2850                 }
2851
2852                 /* This gets us the dx_root */
2853                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2854                 if (ret) {
2855                         mlog_errno(ret);
2856                         goto out;
2857                 }
2858         }
2859
2860         /*
2861          * We should never need more than 2 clusters for the unindexed
2862          * tree - maximum dirent size is far less than one block. In
2863          * fact, the only time we'd need more than one cluster is if
2864          * blocksize == clustersize and the dirent won't fit in the
2865          * extra space that the expansion to a single block gives. As
2866          * of today, that only happens on 4k/4k file systems.
2867          */
2868         BUG_ON(alloc > 2);
2869
2870         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2871         if (ret) {
2872                 mlog_errno(ret);
2873                 goto out;
2874         }
2875
2876         /*
2877          * Prepare for worst case allocation scenario of two separate
2878          * extents in the unindexed tree.
2879          */
2880         if (alloc == 2)
2881                 credits += OCFS2_SUBALLOC_ALLOC;
2882
2883         handle = ocfs2_start_trans(osb, credits);
2884         if (IS_ERR(handle)) {
2885                 ret = PTR_ERR(handle);
2886                 mlog_errno(ret);
2887                 goto out;
2888         }
2889
2890         ret = dquot_alloc_space_nodirty(dir,
2891                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2892         if (ret)
2893                 goto out_commit;
2894         did_quota = 1;
2895
2896         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2897                 /*
2898                  * Allocate our index cluster first, to maximize the
2899                  * possibility that unindexed leaves grow
2900                  * contiguously.
2901                  */
2902                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2903                                                  dx_leaves, num_dx_leaves,
2904                                                  &dx_insert_blkno);
2905                 if (ret) {
2906                         mlog_errno(ret);
2907                         goto out_commit;
2908                 }
2909                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2910         }
2911
2912         /*
2913          * Try to claim as many clusters as the bitmap can give though
2914          * if we only get one now, that's enough to continue. The rest
2915          * will be claimed after the conversion to extents.
2916          */
2917         if (ocfs2_dir_resv_allowed(osb))
2918                 data_ac->ac_resv = &oi->ip_la_data_resv;
2919         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2920         if (ret) {
2921                 mlog_errno(ret);
2922                 goto out_commit;
2923         }
2924         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2925
2926         /*
2927          * Operations are carefully ordered so that we set up the new
2928          * data block first. The conversion from inline data to
2929          * extents follows.
2930          */
2931         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2932         dirdata_bh = sb_getblk(sb, blkno);
2933         if (!dirdata_bh) {
2934                 ret = -ENOMEM;
2935                 mlog_errno(ret);
2936                 goto out_commit;
2937         }
2938
2939         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2940
2941         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2942                                       OCFS2_JOURNAL_ACCESS_CREATE);
2943         if (ret) {
2944                 mlog_errno(ret);
2945                 goto out_commit;
2946         }
2947
2948         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2949         memset(dirdata_bh->b_data + i_size_read(dir), 0,
2950                sb->s_blocksize - i_size_read(dir));
2951         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2952         if (ocfs2_new_dir_wants_trailer(dir)) {
2953                 /*
2954                  * Prepare the dir trailer up front. It will otherwise look
2955                  * like a valid dirent. Even if inserting the index fails
2956                  * (unlikely), then all we'll have done is given first dir
2957                  * block a small amount of fragmentation.
2958                  */
2959                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2960         }
2961
2962         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2963         ocfs2_journal_dirty(handle, dirdata_bh);
2964
2965         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2966                 /*
2967                  * Dx dirs with an external cluster need to do this up
2968                  * front. Inline dx root's get handled later, after
2969                  * we've allocated our root block. We get passed back
2970                  * a total number of items so that dr_num_entries can
2971                  * be correctly set once the dx_root has been
2972                  * allocated.
2973                  */
2974                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2975                                                num_dx_leaves, &num_dx_entries,
2976                                                dirdata_bh);
2977                 if (ret) {
2978                         mlog_errno(ret);
2979                         goto out_commit;
2980                 }
2981         }
2982
2983         /*
2984          * Set extent, i_size, etc on the directory. After this, the
2985          * inode should contain the same exact dirents as before and
2986          * be fully accessible from system calls.
2987          *
2988          * We let the later dirent insert modify c/mtime - to the user
2989          * the data hasn't changed.
2990          */
2991         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2992                                       OCFS2_JOURNAL_ACCESS_CREATE);
2993         if (ret) {
2994                 mlog_errno(ret);
2995                 goto out_commit;
2996         }
2997
2998         spin_lock(&oi->ip_lock);
2999         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
3000         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
3001         spin_unlock(&oi->ip_lock);
3002
3003         ocfs2_dinode_new_extent_list(dir, di);
3004
3005         i_size_write(dir, sb->s_blocksize);
3006         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3007
3008         di->i_size = cpu_to_le64(sb->s_blocksize);
3009         di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3010         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3011         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3012
3013         /*
3014          * This should never fail as our extent list is empty and all
3015          * related blocks have been journaled already.
3016          */
3017         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
3018                                   0, NULL);
3019         if (ret) {
3020                 mlog_errno(ret);
3021                 goto out_commit;
3022         }
3023
3024         /*
3025          * Set i_blocks after the extent insert for the most up to
3026          * date ip_clusters value.
3027          */
3028         dir->i_blocks = ocfs2_inode_sector_count(dir);
3029
3030         ocfs2_journal_dirty(handle, di_bh);
3031
3032         if (ocfs2_supports_indexed_dirs(osb)) {
3033                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3034                                                 dirdata_bh, meta_ac, dx_inline,
3035                                                 num_dx_entries, &dx_root_bh);
3036                 if (ret) {
3037                         mlog_errno(ret);
3038                         goto out_commit;
3039                 }
3040
3041                 if (dx_inline) {
3042                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3043                                                       dirdata_bh);
3044                 } else {
3045                         ocfs2_init_dx_root_extent_tree(&dx_et,
3046                                                        INODE_CACHE(dir),
3047                                                        dx_root_bh);
3048                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3049                                                   dx_insert_blkno, 1, 0, NULL);
3050                         if (ret)
3051                                 mlog_errno(ret);
3052                 }
3053         }
3054
3055         /*
3056          * We asked for two clusters, but only got one in the 1st
3057          * pass. Claim the 2nd cluster as a separate extent.
3058          */
3059         if (alloc > len) {
3060                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3061                                            &len);
3062                 if (ret) {
3063                         mlog_errno(ret);
3064                         goto out_commit;
3065                 }
3066                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3067
3068                 ret = ocfs2_insert_extent(handle, &et, 1,
3069                                           blkno, len, 0, NULL);
3070                 if (ret) {
3071                         mlog_errno(ret);
3072                         goto out_commit;
3073                 }
3074                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3075         }
3076
3077         *first_block_bh = dirdata_bh;
3078         dirdata_bh = NULL;
3079         if (ocfs2_supports_indexed_dirs(osb)) {
3080                 unsigned int off;
3081
3082                 if (!dx_inline) {
3083                         /*
3084                          * We need to return the correct block within the
3085                          * cluster which should hold our entry.
3086                          */
3087                         off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3088                                                     &lookup->dl_hinfo);
3089                         get_bh(dx_leaves[off]);
3090                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3091                 }
3092                 lookup->dl_dx_root_bh = dx_root_bh;
3093                 dx_root_bh = NULL;
3094         }
3095
3096 out_commit:
3097         if (ret < 0 && did_quota)
3098                 dquot_free_space_nodirty(dir, bytes_allocated);
3099
3100         ocfs2_commit_trans(osb, handle);
3101
3102 out:
3103         up_write(&oi->ip_alloc_sem);
3104         if (data_ac)
3105                 ocfs2_free_alloc_context(data_ac);
3106         if (meta_ac)
3107                 ocfs2_free_alloc_context(meta_ac);
3108
3109         if (dx_leaves) {
3110                 for (i = 0; i < num_dx_leaves; i++)
3111                         brelse(dx_leaves[i]);
3112                 kfree(dx_leaves);
3113         }
3114
3115         brelse(dirdata_bh);
3116         brelse(dx_root_bh);
3117
3118         return ret;
3119 }
3120
3121 /* returns a bh of the 1st new block in the allocation. */
3122 static int ocfs2_do_extend_dir(struct super_block *sb,
3123                                handle_t *handle,
3124                                struct inode *dir,
3125                                struct buffer_head *parent_fe_bh,
3126                                struct ocfs2_alloc_context *data_ac,
3127                                struct ocfs2_alloc_context *meta_ac,
3128                                struct buffer_head **new_bh)
3129 {
3130         int status;
3131         int extend, did_quota = 0;
3132         u64 p_blkno, v_blkno;
3133
3134         spin_lock(&OCFS2_I(dir)->ip_lock);
3135         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3136         spin_unlock(&OCFS2_I(dir)->ip_lock);
3137
3138         if (extend) {
3139                 u32 offset = OCFS2_I(dir)->ip_clusters;
3140
3141                 status = dquot_alloc_space_nodirty(dir,
3142                                         ocfs2_clusters_to_bytes(sb, 1));
3143                 if (status)
3144                         goto bail;
3145                 did_quota = 1;
3146
3147                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3148                                               1, 0, parent_fe_bh, handle,
3149                                               data_ac, meta_ac, NULL);
3150                 BUG_ON(status == -EAGAIN);
3151                 if (status < 0) {
3152                         mlog_errno(status);
3153                         goto bail;
3154                 }
3155         }
3156
3157         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3158         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3159         if (status < 0) {
3160                 mlog_errno(status);
3161                 goto bail;
3162         }
3163
3164         *new_bh = sb_getblk(sb, p_blkno);
3165         if (!*new_bh) {
3166                 status = -ENOMEM;
3167                 mlog_errno(status);
3168                 goto bail;
3169         }
3170         status = 0;
3171 bail:
3172         if (did_quota && status < 0)
3173                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3174         return status;
3175 }
3176
3177 /*
3178  * Assumes you already have a cluster lock on the directory.
3179  *
3180  * 'blocks_wanted' is only used if we have an inline directory which
3181  * is to be turned into an extent based one. The size of the dirent to
3182  * insert might be larger than the space gained by growing to just one
3183  * block, so we may have to grow the inode by two blocks in that case.
3184  *
3185  * If the directory is already indexed, dx_root_bh must be provided.
3186  */
3187 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3188                             struct inode *dir,
3189                             struct buffer_head *parent_fe_bh,
3190                             unsigned int blocks_wanted,
3191                             struct ocfs2_dir_lookup_result *lookup,
3192                             struct buffer_head **new_de_bh)
3193 {
3194         int status = 0;
3195         int credits, num_free_extents, drop_alloc_sem = 0;
3196         loff_t dir_i_size;
3197         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3198         struct ocfs2_extent_list *el = &fe->id2.i_list;
3199         struct ocfs2_alloc_context *data_ac = NULL;
3200         struct ocfs2_alloc_context *meta_ac = NULL;
3201         handle_t *handle = NULL;
3202         struct buffer_head *new_bh = NULL;
3203         struct ocfs2_dir_entry * de;
3204         struct super_block *sb = osb->sb;
3205         struct ocfs2_extent_tree et;
3206         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3207
3208         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3209                 /*
3210                  * This would be a code error as an inline directory should
3211                  * never have an index root.
3212                  */
3213                 BUG_ON(dx_root_bh);
3214
3215                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3216                                                  blocks_wanted, lookup,
3217                                                  &new_bh);
3218                 if (status) {
3219                         mlog_errno(status);
3220                         goto bail;
3221                 }
3222
3223                 /* Expansion from inline to an indexed directory will
3224                  * have given us this. */
3225                 dx_root_bh = lookup->dl_dx_root_bh;
3226
3227                 if (blocks_wanted == 1) {
3228                         /*
3229                          * If the new dirent will fit inside the space
3230                          * created by pushing out to one block, then
3231                          * we can complete the operation
3232                          * here. Otherwise we have to expand i_size
3233                          * and format the 2nd block below.
3234                          */
3235                         BUG_ON(new_bh == NULL);
3236                         goto bail_bh;
3237                 }
3238
3239                 /*
3240                  * Get rid of 'new_bh' - we want to format the 2nd
3241                  * data block and return that instead.
3242                  */
3243                 brelse(new_bh);
3244                 new_bh = NULL;
3245
3246                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3247                 drop_alloc_sem = 1;
3248                 dir_i_size = i_size_read(dir);
3249                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3250                 goto do_extend;
3251         }
3252
3253         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3254         drop_alloc_sem = 1;
3255         dir_i_size = i_size_read(dir);
3256         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3257                                dir_i_size);
3258
3259         /* dir->i_size is always block aligned. */
3260         spin_lock(&OCFS2_I(dir)->ip_lock);
3261         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3262                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3263                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3264                                               parent_fe_bh);
3265                 num_free_extents = ocfs2_num_free_extents(osb, &et);
3266                 if (num_free_extents < 0) {
3267                         status = num_free_extents;
3268                         mlog_errno(status);
3269                         goto bail;
3270                 }
3271
3272                 if (!num_free_extents) {
3273                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3274                         if (status < 0) {
3275                                 if (status != -ENOSPC)
3276                                         mlog_errno(status);
3277                                 goto bail;
3278                         }
3279                 }
3280
3281                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3282                 if (status < 0) {
3283                         if (status != -ENOSPC)
3284                                 mlog_errno(status);
3285                         goto bail;
3286                 }
3287
3288                 if (ocfs2_dir_resv_allowed(osb))
3289                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3290
3291                 credits = ocfs2_calc_extend_credits(sb, el);
3292         } else {
3293                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3294                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3295         }
3296
3297 do_extend:
3298         if (ocfs2_dir_indexed(dir))
3299                 credits++; /* For attaching the new dirent block to the
3300                             * dx_root */
3301
3302         handle = ocfs2_start_trans(osb, credits);
3303         if (IS_ERR(handle)) {
3304                 status = PTR_ERR(handle);
3305                 handle = NULL;
3306                 mlog_errno(status);
3307                 goto bail;
3308         }
3309
3310         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3311                                      data_ac, meta_ac, &new_bh);
3312         if (status < 0) {
3313                 mlog_errno(status);
3314                 goto bail;
3315         }
3316
3317         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3318
3319         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3320                                          OCFS2_JOURNAL_ACCESS_CREATE);
3321         if (status < 0) {
3322                 mlog_errno(status);
3323                 goto bail;
3324         }
3325         memset(new_bh->b_data, 0, sb->s_blocksize);
3326
3327         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3328         de->inode = 0;
3329         if (ocfs2_supports_dir_trailer(dir)) {
3330                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3331
3332                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3333
3334                 if (ocfs2_dir_indexed(dir)) {
3335                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3336                                                            dx_root_bh, new_bh);
3337                         if (status) {
3338                                 mlog_errno(status);
3339                                 goto bail;
3340                         }
3341                 }
3342         } else {
3343                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3344         }
3345         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3346         ocfs2_journal_dirty(handle, new_bh);
3347
3348         dir_i_size += dir->i_sb->s_blocksize;
3349         i_size_write(dir, dir_i_size);
3350         dir->i_blocks = ocfs2_inode_sector_count(dir);
3351         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3352         if (status < 0) {
3353                 mlog_errno(status);
3354                 goto bail;
3355         }
3356
3357 bail_bh:
3358         *new_de_bh = new_bh;
3359         get_bh(*new_de_bh);
3360 bail:
3361         if (handle)
3362                 ocfs2_commit_trans(osb, handle);
3363         if (drop_alloc_sem)
3364                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3365
3366         if (data_ac)
3367                 ocfs2_free_alloc_context(data_ac);
3368         if (meta_ac)
3369                 ocfs2_free_alloc_context(meta_ac);
3370
3371         brelse(new_bh);
3372
3373         return status;
3374 }
3375
3376 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3377                                    const char *name, int namelen,
3378                                    struct buffer_head **ret_de_bh,
3379                                    unsigned int *blocks_wanted)
3380 {
3381         int ret;
3382         struct super_block *sb = dir->i_sb;
3383         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3384         struct ocfs2_dir_entry *de, *last_de = NULL;
3385         char *de_buf, *limit;
3386         unsigned long offset = 0;
3387         unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3388
3389         /*
3390          * This calculates how many free bytes we'd have in block zero, should
3391          * this function force expansion to an extent tree.
3392          */
3393         if (ocfs2_new_dir_wants_trailer(dir))
3394                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3395         else
3396                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3397
3398         de_buf = di->id2.i_data.id_data;
3399         limit = de_buf + i_size_read(dir);
3400         rec_len = OCFS2_DIR_REC_LEN(namelen);
3401
3402         while (de_buf < limit) {
3403                 de = (struct ocfs2_dir_entry *)de_buf;
3404
3405                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3406                         ret = -ENOENT;
3407                         goto out;
3408                 }
3409                 if (ocfs2_match(namelen, name, de)) {
3410                         ret = -EEXIST;
3411                         goto out;
3412                 }
3413                 /*
3414                  * No need to check for a trailing dirent record here as
3415                  * they're not used for inline dirs.
3416                  */
3417
3418                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3419                         /* Ok, we found a spot. Return this bh and let
3420                          * the caller actually fill it in. */
3421                         *ret_de_bh = di_bh;
3422                         get_bh(*ret_de_bh);
3423                         ret = 0;
3424                         goto out;
3425                 }
3426
3427                 last_de = de;
3428                 de_buf += le16_to_cpu(de->rec_len);
3429                 offset += le16_to_cpu(de->rec_len);
3430         }
3431
3432         /*
3433          * We're going to require expansion of the directory - figure
3434          * out how many blocks we'll need so that a place for the
3435          * dirent can be found.
3436          */
3437         *blocks_wanted = 1;
3438         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3439         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3440                 *blocks_wanted = 2;
3441
3442         ret = -ENOSPC;
3443 out:
3444         return ret;
3445 }
3446
3447 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3448                                    int namelen, struct buffer_head **ret_de_bh)
3449 {
3450         unsigned long offset;
3451         struct buffer_head *bh = NULL;
3452         unsigned short rec_len;
3453         struct ocfs2_dir_entry *de;
3454         struct super_block *sb = dir->i_sb;
3455         int status;
3456         int blocksize = dir->i_sb->s_blocksize;
3457
3458         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3459         if (status) {
3460                 mlog_errno(status);
3461                 goto bail;
3462         }
3463
3464         rec_len = OCFS2_DIR_REC_LEN(namelen);
3465         offset = 0;
3466         de = (struct ocfs2_dir_entry *) bh->b_data;
3467         while (1) {
3468                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3469                         brelse(bh);
3470                         bh = NULL;
3471
3472                         if (i_size_read(dir) <= offset) {
3473                                 /*
3474                                  * Caller will have to expand this
3475                                  * directory.
3476                                  */
3477                                 status = -ENOSPC;
3478                                 goto bail;
3479                         }
3480                         status = ocfs2_read_dir_block(dir,
3481                                              offset >> sb->s_blocksize_bits,
3482                                              &bh, 0);
3483                         if (status) {
3484                                 mlog_errno(status);
3485                                 goto bail;
3486                         }
3487                         /* move to next block */
3488                         de = (struct ocfs2_dir_entry *) bh->b_data;
3489                 }
3490                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3491                         status = -ENOENT;
3492                         goto bail;
3493                 }
3494                 if (ocfs2_match(namelen, name, de)) {
3495                         status = -EEXIST;
3496                         goto bail;
3497                 }
3498
3499                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3500                                            blocksize))
3501                         goto next;
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 = bh;
3507                         get_bh(*ret_de_bh);
3508                         status = 0;
3509                         goto bail;
3510                 }
3511 next:
3512                 offset += le16_to_cpu(de->rec_len);
3513                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3514         }
3515
3516         status = 0;
3517 bail:
3518         brelse(bh);
3519         if (status)
3520                 mlog_errno(status);
3521
3522         return status;
3523 }
3524
3525 static int dx_leaf_sort_cmp(const void *a, const void *b)
3526 {
3527         const struct ocfs2_dx_entry *entry1 = a;
3528         const struct ocfs2_dx_entry *entry2 = b;
3529         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3530         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3531         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3532         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3533
3534         if (major_hash1 > major_hash2)
3535                 return 1;
3536         if (major_hash1 < major_hash2)
3537                 return -1;
3538
3539         /*
3540          * It is not strictly necessary to sort by minor
3541          */
3542         if (minor_hash1 > minor_hash2)
3543                 return 1;
3544         if (minor_hash1 < minor_hash2)
3545                 return -1;
3546         return 0;
3547 }
3548
3549 static void dx_leaf_sort_swap(void *a, void *b, int size)
3550 {
3551         struct ocfs2_dx_entry *entry1 = a;
3552         struct ocfs2_dx_entry *entry2 = b;
3553         struct ocfs2_dx_entry tmp;
3554
3555         BUG_ON(size != sizeof(*entry1));
3556
3557         tmp = *entry1;
3558         *entry1 = *entry2;
3559         *entry2 = tmp;
3560 }
3561
3562 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3563 {
3564         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3565         int i, num = le16_to_cpu(dl_list->de_num_used);
3566
3567         for (i = 0; i < (num - 1); i++) {
3568                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3569                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3570                         return 0;
3571         }
3572
3573         return 1;
3574 }
3575
3576 /*
3577  * Find the optimal value to split this leaf on. This expects the leaf
3578  * entries to be in sorted order.
3579  *
3580  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3581  * the hash we want to insert.
3582  *
3583  * This function is only concerned with the major hash - that which
3584  * determines which cluster an item belongs to.
3585  */
3586 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3587                                         u32 leaf_cpos, u32 insert_hash,
3588                                         u32 *split_hash)
3589 {
3590         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3591         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3592         int allsame;
3593
3594         /*
3595          * There's a couple rare, but nasty corner cases we have to
3596          * check for here. All of them involve a leaf where all value
3597          * have the same hash, which is what we look for first.
3598          *
3599          * Most of the time, all of the above is false, and we simply
3600          * pick the median value for a split.
3601          */
3602         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3603         if (allsame) {
3604                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3605
3606                 if (val == insert_hash) {
3607                         /*
3608                          * No matter where we would choose to split,
3609                          * the new entry would want to occupy the same
3610                          * block as these. Since there's no space left
3611                          * in their existing block, we know there
3612                          * won't be space after the split.
3613                          */
3614                         return -ENOSPC;
3615                 }
3616
3617                 if (val == leaf_cpos) {
3618                         /*
3619                          * Because val is the same as leaf_cpos (which
3620                          * is the smallest value this leaf can have),
3621                          * yet is not equal to insert_hash, then we
3622                          * know that insert_hash *must* be larger than
3623                          * val (and leaf_cpos). At least cpos+1 in value.
3624                          *
3625                          * We also know then, that there cannot be an
3626                          * adjacent extent (otherwise we'd be looking
3627                          * at it). Choosing this value gives us a
3628                          * chance to get some contiguousness.
3629                          */
3630                         *split_hash = leaf_cpos + 1;
3631                         return 0;
3632                 }
3633
3634                 if (val > insert_hash) {
3635                         /*
3636                          * val can not be the same as insert hash, and
3637                          * also must be larger than leaf_cpos. Also,
3638                          * we know that there can't be a leaf between
3639                          * cpos and val, otherwise the entries with
3640                          * hash 'val' would be there.
3641                          */
3642                         *split_hash = val;
3643                         return 0;
3644                 }
3645
3646                 *split_hash = insert_hash;
3647                 return 0;
3648         }
3649
3650         /*
3651          * Since the records are sorted and the checks above
3652          * guaranteed that not all records in this block are the same,
3653          * we simple travel forward, from the median, and pick the 1st
3654          * record whose value is larger than leaf_cpos.
3655          */
3656         for (i = (num_used / 2); i < num_used; i++)
3657                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3658                     leaf_cpos)
3659                         break;
3660
3661         BUG_ON(i == num_used); /* Should be impossible */
3662         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3663         return 0;
3664 }
3665
3666 /*
3667  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3668  * larger than split_hash into new_dx_leaves. We use a temporary
3669  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3670  *
3671  * Since the block offset inside a leaf (cluster) is a constant mask
3672  * of minor_hash, we can optimize - an item at block offset X within
3673  * the original cluster, will be at offset X within the new cluster.
3674  */
3675 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3676                                        handle_t *handle,
3677                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3678                                        struct buffer_head **orig_dx_leaves,
3679                                        struct buffer_head **new_dx_leaves,
3680                                        int num_dx_leaves)
3681 {
3682         int i, j, num_used;
3683         u32 major_hash;
3684         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3685         struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3686         struct ocfs2_dx_entry *dx_entry;
3687
3688         tmp_list = &tmp_dx_leaf->dl_list;
3689
3690         for (i = 0; i < num_dx_leaves; i++) {
3691                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3692                 orig_list = &orig_dx_leaf->dl_list;
3693                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3694                 new_list = &new_dx_leaf->dl_list;
3695
3696                 num_used = le16_to_cpu(orig_list->de_num_used);
3697
3698                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3699                 tmp_list->de_num_used = cpu_to_le16(0);
3700                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3701
3702                 for (j = 0; j < num_used; j++) {
3703                         dx_entry = &orig_list->de_entries[j];
3704                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3705                         if (major_hash >= split_hash)
3706                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3707                                                               dx_entry);
3708                         else
3709                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3710                                                               dx_entry);
3711                 }
3712                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3713
3714                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3715                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3716         }
3717 }
3718
3719 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3720                                           struct ocfs2_dx_root_block *dx_root)
3721 {
3722         int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3723
3724         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3725         credits += ocfs2_quota_trans_credits(osb->sb);
3726         return credits;
3727 }
3728
3729 /*
3730  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3731  * half our entries into.
3732  */
3733 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3734                                   struct buffer_head *dx_root_bh,
3735                                   struct buffer_head *dx_leaf_bh,
3736                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3737                                   u64 leaf_blkno)
3738 {
3739         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3740         int credits, ret, i, num_used, did_quota = 0;
3741         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3742         u64 orig_leaves_start;
3743         int num_dx_leaves;
3744         struct buffer_head **orig_dx_leaves = NULL;
3745         struct buffer_head **new_dx_leaves = NULL;
3746         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3747         struct ocfs2_extent_tree et;
3748         handle_t *handle = NULL;
3749         struct ocfs2_dx_root_block *dx_root;
3750         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3751
3752         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3753                                      (unsigned long long)leaf_blkno,
3754                                      insert_hash);
3755
3756         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3757
3758         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3759         /*
3760          * XXX: This is a rather large limit. We should use a more
3761          * realistic value.
3762          */
3763         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3764                 return -ENOSPC;
3765
3766         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3767         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3768                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3769                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3770                      (unsigned long long)leaf_blkno, num_used);
3771                 ret = -EIO;
3772                 goto out;
3773         }
3774
3775         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3776         if (!orig_dx_leaves) {
3777                 ret = -ENOMEM;
3778                 mlog_errno(ret);
3779                 goto out;
3780         }
3781
3782         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3783         if (!new_dx_leaves) {
3784                 ret = -ENOMEM;
3785                 mlog_errno(ret);
3786                 goto out;
3787         }
3788
3789         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3790         if (ret) {
3791                 if (ret != -ENOSPC)
3792                         mlog_errno(ret);
3793                 goto out;
3794         }
3795
3796         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3797         handle = ocfs2_start_trans(osb, credits);
3798         if (IS_ERR(handle)) {
3799                 ret = PTR_ERR(handle);
3800                 handle = NULL;
3801                 mlog_errno(ret);
3802                 goto out;
3803         }
3804
3805         ret = dquot_alloc_space_nodirty(dir,
3806                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3807         if (ret)
3808                 goto out_commit;
3809         did_quota = 1;
3810
3811         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3812                                       OCFS2_JOURNAL_ACCESS_WRITE);
3813         if (ret) {
3814                 mlog_errno(ret);
3815                 goto out_commit;
3816         }
3817
3818         /*
3819          * This block is changing anyway, so we can sort it in place.
3820          */
3821         sort(dx_leaf->dl_list.de_entries, num_used,
3822              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3823              dx_leaf_sort_swap);
3824
3825         ocfs2_journal_dirty(handle, dx_leaf_bh);
3826
3827         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3828                                            &split_hash);
3829         if (ret) {
3830                 mlog_errno(ret);
3831                 goto  out_commit;
3832         }
3833
3834         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3835
3836         /*
3837          * We have to carefully order operations here. There are items
3838          * which want to be in the new cluster before insert, but in
3839          * order to put those items in the new cluster, we alter the
3840          * old cluster. A failure to insert gets nasty.
3841          *
3842          * So, start by reserving writes to the old
3843          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3844          * the new cluster for us, before inserting it. The insert
3845          * won't happen if there's an error before that. Once the
3846          * insert is done then, we can transfer from one leaf into the
3847          * other without fear of hitting any error.
3848          */
3849
3850         /*
3851          * The leaf transfer wants some scratch space so that we don't
3852          * wind up doing a bunch of expensive memmove().
3853          */
3854         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3855         if (!tmp_dx_leaf) {
3856                 ret = -ENOMEM;
3857                 mlog_errno(ret);
3858                 goto out_commit;
3859         }
3860
3861         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3862         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3863                                    orig_dx_leaves);
3864         if (ret) {
3865                 mlog_errno(ret);
3866                 goto out_commit;
3867         }
3868
3869         cpos = split_hash;
3870         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3871                                        data_ac, meta_ac, new_dx_leaves,
3872                                        num_dx_leaves);
3873         if (ret) {
3874                 mlog_errno(ret);
3875                 goto out_commit;
3876         }
3877
3878         for (i = 0; i < num_dx_leaves; i++) {
3879                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3880                                               orig_dx_leaves[i],
3881                                               OCFS2_JOURNAL_ACCESS_WRITE);
3882                 if (ret) {
3883                         mlog_errno(ret);
3884                         goto out_commit;
3885                 }
3886
3887                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3888                                               new_dx_leaves[i],
3889                                               OCFS2_JOURNAL_ACCESS_WRITE);
3890                 if (ret) {
3891                         mlog_errno(ret);
3892                         goto out_commit;
3893                 }
3894         }
3895
3896         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3897                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3898
3899 out_commit:
3900         if (ret < 0 && did_quota)
3901                 dquot_free_space_nodirty(dir,
3902                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3903
3904         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3905         ocfs2_commit_trans(osb, handle);
3906
3907 out:
3908         if (orig_dx_leaves || new_dx_leaves) {
3909                 for (i = 0; i < num_dx_leaves; i++) {
3910                         if (orig_dx_leaves)
3911                                 brelse(orig_dx_leaves[i]);
3912                         if (new_dx_leaves)
3913                                 brelse(new_dx_leaves[i]);
3914                 }
3915                 kfree(orig_dx_leaves);
3916                 kfree(new_dx_leaves);
3917         }
3918
3919         if (meta_ac)
3920                 ocfs2_free_alloc_context(meta_ac);
3921         if (data_ac)
3922                 ocfs2_free_alloc_context(data_ac);
3923
3924         kfree(tmp_dx_leaf);
3925         return ret;
3926 }
3927
3928 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3929                                    struct buffer_head *di_bh,
3930                                    struct buffer_head *dx_root_bh,
3931                                    const char *name, int namelen,
3932                                    struct ocfs2_dir_lookup_result *lookup)
3933 {
3934         int ret, rebalanced = 0;
3935         struct ocfs2_dx_root_block *dx_root;
3936         struct buffer_head *dx_leaf_bh = NULL;
3937         struct ocfs2_dx_leaf *dx_leaf;
3938         u64 blkno;
3939         u32 leaf_cpos;
3940
3941         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3942
3943 restart_search:
3944         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3945                                   &leaf_cpos, &blkno);
3946         if (ret) {
3947                 mlog_errno(ret);
3948                 goto out;
3949         }
3950
3951         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3952         if (ret) {
3953                 mlog_errno(ret);
3954                 goto out;
3955         }
3956
3957         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3958
3959         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3960             le16_to_cpu(dx_leaf->dl_list.de_count)) {
3961                 if (rebalanced) {
3962                         /*
3963                          * Rebalancing should have provided us with
3964                          * space in an appropriate leaf.
3965                          *
3966                          * XXX: Is this an abnormal condition then?
3967                          * Should we print a message here?
3968                          */
3969                         ret = -ENOSPC;
3970                         goto out;
3971                 }
3972
3973                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3974                                              &lookup->dl_hinfo, leaf_cpos,
3975                                              blkno);
3976                 if (ret) {
3977                         if (ret != -ENOSPC)
3978                                 mlog_errno(ret);
3979                         goto out;
3980                 }
3981
3982                 /*
3983                  * Restart the lookup. The rebalance might have
3984                  * changed which block our item fits into. Mark our
3985                  * progress, so we only execute this once.
3986                  */
3987                 brelse(dx_leaf_bh);
3988                 dx_leaf_bh = NULL;
3989                 rebalanced = 1;
3990                 goto restart_search;
3991         }
3992
3993         lookup->dl_dx_leaf_bh = dx_leaf_bh;
3994         dx_leaf_bh = NULL;
3995
3996 out:
3997         brelse(dx_leaf_bh);
3998         return ret;
3999 }
4000
4001 static int ocfs2_search_dx_free_list(struct inode *dir,
4002                                      struct buffer_head *dx_root_bh,
4003                                      int namelen,
4004                                      struct ocfs2_dir_lookup_result *lookup)
4005 {
4006         int ret = -ENOSPC;
4007         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4008         struct ocfs2_dir_block_trailer *db;
4009         u64 next_block;
4010         int rec_len = OCFS2_DIR_REC_LEN(namelen);
4011         struct ocfs2_dx_root_block *dx_root;
4012
4013         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4014         next_block = le64_to_cpu(dx_root->dr_free_blk);
4015
4016         while (next_block) {
4017                 brelse(prev_leaf_bh);
4018                 prev_leaf_bh = leaf_bh;
4019                 leaf_bh = NULL;
4020
4021                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4022                 if (ret) {
4023                         mlog_errno(ret);
4024                         goto out;
4025                 }
4026
4027                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4028                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4029                         lookup->dl_leaf_bh = leaf_bh;
4030                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
4031                         leaf_bh = NULL;
4032                         prev_leaf_bh = NULL;
4033                         break;
4034                 }
4035
4036                 next_block = le64_to_cpu(db->db_free_next);
4037         }
4038
4039         if (!next_block)
4040                 ret = -ENOSPC;
4041
4042 out:
4043
4044         brelse(leaf_bh);
4045         brelse(prev_leaf_bh);
4046         return ret;
4047 }
4048
4049 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4050                                        struct buffer_head *dx_root_bh)
4051 {
4052         int ret, num_dx_leaves, i, j, did_quota = 0;
4053         struct buffer_head **dx_leaves = NULL;
4054         struct ocfs2_extent_tree et;
4055         u64 insert_blkno;
4056         struct ocfs2_alloc_context *data_ac = NULL;
4057         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4058         handle_t *handle = NULL;
4059         struct ocfs2_dx_root_block *dx_root;
4060         struct ocfs2_dx_entry_list *entry_list;
4061         struct ocfs2_dx_entry *dx_entry;
4062         struct ocfs2_dx_leaf *target_leaf;
4063
4064         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4065         if (ret) {
4066                 mlog_errno(ret);
4067                 goto out;
4068         }
4069
4070         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4071         if (!dx_leaves) {
4072                 ret = -ENOMEM;
4073                 mlog_errno(ret);
4074                 goto out;
4075         }
4076
4077         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4078         if (IS_ERR(handle)) {
4079                 ret = PTR_ERR(handle);
4080                 mlog_errno(ret);
4081                 goto out;
4082         }
4083
4084         ret = dquot_alloc_space_nodirty(dir,
4085                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4086         if (ret)
4087                 goto out_commit;
4088         did_quota = 1;
4089
4090         /*
4091          * We do this up front, before the allocation, so that a
4092          * failure to add the dx_root_bh to the journal won't result
4093          * us losing clusters.
4094          */
4095         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4096                                       OCFS2_JOURNAL_ACCESS_WRITE);
4097         if (ret) {
4098                 mlog_errno(ret);
4099                 goto out_commit;
4100         }
4101
4102         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4103                                          num_dx_leaves, &insert_blkno);
4104         if (ret) {
4105                 mlog_errno(ret);
4106                 goto out_commit;
4107         }
4108
4109         /*
4110          * Transfer the entries from our dx_root into the appropriate
4111          * block
4112          */
4113         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4114         entry_list = &dx_root->dr_entries;
4115
4116         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4117                 dx_entry = &entry_list->de_entries[i];
4118
4119                 j = __ocfs2_dx_dir_hash_idx(osb,
4120                                             le32_to_cpu(dx_entry->dx_minor_hash));
4121                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4122
4123                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4124
4125                 /* Each leaf has been passed to the journal already
4126                  * via __ocfs2_dx_dir_new_cluster() */
4127         }
4128
4129         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4130         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4131                offsetof(struct ocfs2_dx_root_block, dr_list));
4132         dx_root->dr_list.l_count =
4133                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4134
4135         /* This should never fail considering we start with an empty
4136          * dx_root. */
4137         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4138         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4139         if (ret)
4140                 mlog_errno(ret);
4141         did_quota = 0;
4142
4143         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4144         ocfs2_journal_dirty(handle, dx_root_bh);
4145
4146 out_commit:
4147         if (ret < 0 && did_quota)
4148                 dquot_free_space_nodirty(dir,
4149                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4150
4151         ocfs2_commit_trans(osb, handle);
4152
4153 out:
4154         if (data_ac)
4155                 ocfs2_free_alloc_context(data_ac);
4156
4157         if (dx_leaves) {
4158                 for (i = 0; i < num_dx_leaves; i++)
4159                         brelse(dx_leaves[i]);
4160                 kfree(dx_leaves);
4161         }
4162         return ret;
4163 }
4164
4165 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4166 {
4167         struct ocfs2_dx_root_block *dx_root;
4168         struct ocfs2_dx_entry_list *entry_list;
4169
4170         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4171         entry_list = &dx_root->dr_entries;
4172
4173         if (le16_to_cpu(entry_list->de_num_used) >=
4174             le16_to_cpu(entry_list->de_count))
4175                 return -ENOSPC;
4176
4177         return 0;
4178 }
4179
4180 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4181                                            struct buffer_head *di_bh,
4182                                            const char *name,
4183                                            int namelen,
4184                                            struct ocfs2_dir_lookup_result *lookup)
4185 {
4186         int ret, free_dx_root = 1;
4187         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4188         struct buffer_head *dx_root_bh = NULL;
4189         struct buffer_head *leaf_bh = NULL;
4190         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4191         struct ocfs2_dx_root_block *dx_root;
4192
4193         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4194         if (ret) {
4195                 mlog_errno(ret);
4196                 goto out;
4197         }
4198
4199         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4200         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4201                 ret = -ENOSPC;
4202                 mlog_errno(ret);
4203                 goto out;
4204         }
4205
4206         if (ocfs2_dx_root_inline(dx_root)) {
4207                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4208
4209                 if (ret == 0)
4210                         goto search_el;
4211
4212                 /*
4213                  * We ran out of room in the root block. Expand it to
4214                  * an extent, then allow ocfs2_find_dir_space_dx to do
4215                  * the rest.
4216                  */
4217                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4218                 if (ret) {
4219                         mlog_errno(ret);
4220                         goto out;
4221                 }
4222         }
4223
4224         /*
4225          * Insert preparation for an indexed directory is split into two
4226          * steps. The call to find_dir_space_dx reserves room in the index for
4227          * an additional item. If we run out of space there, it's a real error
4228          * we can't continue on.
4229          */
4230         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4231                                       namelen, lookup);
4232         if (ret) {
4233                 mlog_errno(ret);
4234                 goto out;
4235         }
4236
4237 search_el:
4238         /*
4239          * Next, we need to find space in the unindexed tree. This call
4240          * searches using the free space linked list. If the unindexed tree
4241          * lacks sufficient space, we'll expand it below. The expansion code
4242          * is smart enough to add any new blocks to the free space list.
4243          */
4244         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4245         if (ret && ret != -ENOSPC) {
4246                 mlog_errno(ret);
4247                 goto out;
4248         }
4249
4250         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4251         lookup->dl_dx_root_bh = dx_root_bh;
4252         free_dx_root = 0;
4253
4254         if (ret == -ENOSPC) {
4255                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4256
4257                 if (ret) {
4258                         mlog_errno(ret);
4259                         goto out;
4260                 }
4261
4262                 /*
4263                  * We make the assumption here that new leaf blocks are added
4264                  * to the front of our free list.
4265                  */
4266                 lookup->dl_prev_leaf_bh = NULL;
4267                 lookup->dl_leaf_bh = leaf_bh;
4268         }
4269
4270 out:
4271         if (free_dx_root)
4272                 brelse(dx_root_bh);
4273         return ret;
4274 }
4275
4276 /*
4277  * Get a directory ready for insert. Any directory allocation required
4278  * happens here. Success returns zero, and enough context in the dir
4279  * lookup result that ocfs2_add_entry() will be able complete the task
4280  * with minimal performance impact.
4281  */
4282 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4283                                  struct inode *dir,
4284                                  struct buffer_head *parent_fe_bh,
4285                                  const char *name,
4286                                  int namelen,
4287                                  struct ocfs2_dir_lookup_result *lookup)
4288 {
4289         int ret;
4290         unsigned int blocks_wanted = 1;
4291         struct buffer_head *bh = NULL;
4292
4293         trace_ocfs2_prepare_dir_for_insert(
4294                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4295
4296         if (!namelen) {
4297                 ret = -EINVAL;
4298                 mlog_errno(ret);
4299                 goto out;
4300         }
4301
4302         /*
4303          * Do this up front to reduce confusion.
4304          *
4305          * The directory might start inline, then be turned into an
4306          * indexed one, in which case we'd need to hash deep inside
4307          * ocfs2_find_dir_space_id(). Since
4308          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4309          * done, there seems no point in spreading out the calls. We
4310          * can optimize away the case where the file system doesn't
4311          * support indexing.
4312          */
4313         if (ocfs2_supports_indexed_dirs(osb))
4314                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4315
4316         if (ocfs2_dir_indexed(dir)) {
4317                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4318                                                       name, namelen, lookup);
4319                 if (ret)
4320                         mlog_errno(ret);
4321                 goto out;
4322         }
4323
4324         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4325                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4326                                               namelen, &bh, &blocks_wanted);
4327         } else
4328                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4329
4330         if (ret && ret != -ENOSPC) {
4331                 mlog_errno(ret);
4332                 goto out;
4333         }
4334
4335         if (ret == -ENOSPC) {
4336                 /*
4337                  * We have to expand the directory to add this name.
4338                  */
4339                 BUG_ON(bh);
4340
4341                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4342                                        lookup, &bh);
4343                 if (ret) {
4344                         if (ret != -ENOSPC)
4345                                 mlog_errno(ret);
4346                         goto out;
4347                 }
4348
4349                 BUG_ON(!bh);
4350         }
4351
4352         lookup->dl_leaf_bh = bh;
4353         bh = NULL;
4354 out:
4355         brelse(bh);
4356         return ret;
4357 }
4358
4359 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4360                                      struct buffer_head *di_bh,
4361                                      struct buffer_head *dx_root_bh)
4362 {
4363         int ret;
4364         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4365         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4366         struct ocfs2_dx_root_block *dx_root;
4367         struct inode *dx_alloc_inode = NULL;
4368         struct buffer_head *dx_alloc_bh = NULL;
4369         handle_t *handle;
4370         u64 blk;
4371         u16 bit;
4372         u64 bg_blkno;
4373
4374         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4375
4376         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4377                                         EXTENT_ALLOC_SYSTEM_INODE,
4378                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4379         if (!dx_alloc_inode) {
4380                 ret = -ENOMEM;
4381                 mlog_errno(ret);
4382                 goto out;
4383         }
4384         mutex_lock(&dx_alloc_inode->i_mutex);
4385
4386         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4387         if (ret) {
4388                 mlog_errno(ret);
4389                 goto out_mutex;
4390         }
4391
4392         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4393         if (IS_ERR(handle)) {
4394                 ret = PTR_ERR(handle);
4395                 mlog_errno(ret);
4396                 goto out_unlock;
4397         }
4398
4399         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4400                                       OCFS2_JOURNAL_ACCESS_WRITE);
4401         if (ret) {
4402                 mlog_errno(ret);
4403                 goto out_commit;
4404         }
4405
4406         spin_lock(&OCFS2_I(dir)->ip_lock);
4407         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4408         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4409         spin_unlock(&OCFS2_I(dir)->ip_lock);
4410         di->i_dx_root = cpu_to_le64(0ULL);
4411         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4412
4413         ocfs2_journal_dirty(handle, di_bh);
4414
4415         blk = le64_to_cpu(dx_root->dr_blkno);
4416         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4417         if (dx_root->dr_suballoc_loc)
4418                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4419         else
4420                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4421         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4422                                        bit, bg_blkno, 1);
4423         if (ret)
4424                 mlog_errno(ret);
4425
4426 out_commit:
4427         ocfs2_commit_trans(osb, handle);
4428
4429 out_unlock:
4430         ocfs2_inode_unlock(dx_alloc_inode, 1);
4431
4432 out_mutex:
4433         mutex_unlock(&dx_alloc_inode->i_mutex);
4434         brelse(dx_alloc_bh);
4435 out:
4436         iput(dx_alloc_inode);
4437         return ret;
4438 }
4439
4440 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4441 {
4442         int ret;
4443         unsigned int uninitialized_var(clen);
4444         u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4445         u64 uninitialized_var(blkno);
4446         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4447         struct buffer_head *dx_root_bh = NULL;
4448         struct ocfs2_dx_root_block *dx_root;
4449         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4450         struct ocfs2_cached_dealloc_ctxt dealloc;
4451         struct ocfs2_extent_tree et;
4452
4453         ocfs2_init_dealloc_ctxt(&dealloc);
4454
4455         if (!ocfs2_dir_indexed(dir))
4456                 return 0;
4457
4458         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4459         if (ret) {
4460                 mlog_errno(ret);
4461                 goto out;
4462         }
4463         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4464
4465         if (ocfs2_dx_root_inline(dx_root))
4466                 goto remove_index;
4467
4468         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4469
4470         /* XXX: What if dr_clusters is too large? */
4471         while (le32_to_cpu(dx_root->dr_clusters)) {
4472                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4473                                               major_hash, &cpos, &blkno, &clen);
4474                 if (ret) {
4475                         mlog_errno(ret);
4476                         goto out;
4477                 }
4478
4479                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4480
4481                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4482                                                &dealloc, 0, false);
4483                 if (ret) {
4484                         mlog_errno(ret);
4485                         goto out;
4486                 }
4487
4488                 if (cpos == 0)
4489                         break;
4490
4491                 major_hash = cpos - 1;
4492         }
4493
4494 remove_index:
4495         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4496         if (ret) {
4497                 mlog_errno(ret);
4498                 goto out;
4499         }
4500
4501         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4502 out:
4503         ocfs2_schedule_truncate_log_flush(osb, 1);
4504         ocfs2_run_deallocs(osb, &dealloc);
4505
4506         brelse(dx_root_bh);
4507         return ret;
4508 }