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                         ocfs2_error(dir->i_sb, "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(void *priv, const char *name, int name_len,
2077                                    loff_t pos, u64 ino, unsigned type)
2078 {
2079         struct ocfs2_empty_dir_priv *p = priv;
2080
2081         /*
2082          * Check the positions of "." and ".." records to be sure
2083          * they're in the correct place.
2084          *
2085          * Indexed directories don't need to proceed past the first
2086          * two entries, so we end the scan after seeing '..'. Despite
2087          * that, we allow the scan to proceed In the event that we
2088          * have a corrupted indexed directory (no dot or dot dot
2089          * entries). This allows us to double check for existing
2090          * entries which might not have been found in the index.
2091          */
2092         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2093                 p->seen_dot = 1;
2094                 return 0;
2095         }
2096
2097         if (name_len == 2 && !strncmp("..", name, 2) &&
2098             pos == OCFS2_DIR_REC_LEN(1)) {
2099                 p->seen_dot_dot = 1;
2100
2101                 if (p->dx_dir && p->seen_dot)
2102                         return 1;
2103
2104                 return 0;
2105         }
2106
2107         p->seen_other = 1;
2108         return 1;
2109 }
2110
2111 static int ocfs2_empty_dir_dx(struct inode *inode,
2112                               struct ocfs2_empty_dir_priv *priv)
2113 {
2114         int ret;
2115         struct buffer_head *di_bh = NULL;
2116         struct buffer_head *dx_root_bh = NULL;
2117         struct ocfs2_dinode *di;
2118         struct ocfs2_dx_root_block *dx_root;
2119
2120         priv->dx_dir = 1;
2121
2122         ret = ocfs2_read_inode_block(inode, &di_bh);
2123         if (ret) {
2124                 mlog_errno(ret);
2125                 goto out;
2126         }
2127         di = (struct ocfs2_dinode *)di_bh->b_data;
2128
2129         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2130         if (ret) {
2131                 mlog_errno(ret);
2132                 goto out;
2133         }
2134         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2135
2136         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2137                 priv->seen_other = 1;
2138
2139 out:
2140         brelse(di_bh);
2141         brelse(dx_root_bh);
2142         return ret;
2143 }
2144
2145 /*
2146  * routine to check that the specified directory is empty (for rmdir)
2147  *
2148  * Returns 1 if dir is empty, zero otherwise.
2149  *
2150  * XXX: This is a performance problem for unindexed directories.
2151  */
2152 int ocfs2_empty_dir(struct inode *inode)
2153 {
2154         int ret;
2155         struct ocfs2_empty_dir_priv priv = {
2156                 .ctx.actor = ocfs2_empty_dir_filldir,
2157         };
2158
2159         if (ocfs2_dir_indexed(inode)) {
2160                 ret = ocfs2_empty_dir_dx(inode, &priv);
2161                 if (ret)
2162                         mlog_errno(ret);
2163                 /*
2164                  * We still run ocfs2_dir_foreach to get the checks
2165                  * for "." and "..".
2166                  */
2167         }
2168
2169         ret = ocfs2_dir_foreach(inode, &priv.ctx);
2170         if (ret)
2171                 mlog_errno(ret);
2172
2173         if (!priv.seen_dot || !priv.seen_dot_dot) {
2174                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2175                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2176                 /*
2177                  * XXX: Is it really safe to allow an unlink to continue?
2178                  */
2179                 return 1;
2180         }
2181
2182         return !priv.seen_other;
2183 }
2184
2185 /*
2186  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2187  * "..", which might be used during creation of a directory with a trailing
2188  * header. It is otherwise safe to ignore the return code.
2189  */
2190 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2191                                                           struct inode *parent,
2192                                                           char *start,
2193                                                           unsigned int size)
2194 {
2195         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2196
2197         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2198         de->name_len = 1;
2199         de->rec_len =
2200                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2201         strcpy(de->name, ".");
2202         ocfs2_set_de_type(de, S_IFDIR);
2203
2204         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2205         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2206         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2207         de->name_len = 2;
2208         strcpy(de->name, "..");
2209         ocfs2_set_de_type(de, S_IFDIR);
2210
2211         return de;
2212 }
2213
2214 /*
2215  * This works together with code in ocfs2_mknod_locked() which sets
2216  * the inline-data flag and initializes the inline-data section.
2217  */
2218 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2219                                  handle_t *handle,
2220                                  struct inode *parent,
2221                                  struct inode *inode,
2222                                  struct buffer_head *di_bh)
2223 {
2224         int ret;
2225         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2226         struct ocfs2_inline_data *data = &di->id2.i_data;
2227         unsigned int size = le16_to_cpu(data->id_count);
2228
2229         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2230                                       OCFS2_JOURNAL_ACCESS_WRITE);
2231         if (ret) {
2232                 mlog_errno(ret);
2233                 goto out;
2234         }
2235
2236         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2237         ocfs2_journal_dirty(handle, di_bh);
2238
2239         i_size_write(inode, size);
2240         set_nlink(inode, 2);
2241         inode->i_blocks = ocfs2_inode_sector_count(inode);
2242
2243         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2244         if (ret < 0)
2245                 mlog_errno(ret);
2246
2247 out:
2248         return ret;
2249 }
2250
2251 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2252                                  handle_t *handle,
2253                                  struct inode *parent,
2254                                  struct inode *inode,
2255                                  struct buffer_head *fe_bh,
2256                                  struct ocfs2_alloc_context *data_ac,
2257                                  struct buffer_head **ret_new_bh)
2258 {
2259         int status;
2260         unsigned int size = osb->sb->s_blocksize;
2261         struct buffer_head *new_bh = NULL;
2262         struct ocfs2_dir_entry *de;
2263
2264         if (ocfs2_new_dir_wants_trailer(inode))
2265                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2266
2267         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2268                                      data_ac, NULL, &new_bh);
2269         if (status < 0) {
2270                 mlog_errno(status);
2271                 goto bail;
2272         }
2273
2274         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2275
2276         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2277                                          OCFS2_JOURNAL_ACCESS_CREATE);
2278         if (status < 0) {
2279                 mlog_errno(status);
2280                 goto bail;
2281         }
2282         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2283
2284         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2285         if (ocfs2_new_dir_wants_trailer(inode)) {
2286                 int size = le16_to_cpu(de->rec_len);
2287
2288                 /*
2289                  * Figure out the size of the hole left over after
2290                  * insertion of '.' and '..'. The trailer wants this
2291                  * information.
2292                  */
2293                 size -= OCFS2_DIR_REC_LEN(2);
2294                 size -= sizeof(struct ocfs2_dir_block_trailer);
2295
2296                 ocfs2_init_dir_trailer(inode, new_bh, size);
2297         }
2298
2299         ocfs2_journal_dirty(handle, new_bh);
2300
2301         i_size_write(inode, inode->i_sb->s_blocksize);
2302         set_nlink(inode, 2);
2303         inode->i_blocks = ocfs2_inode_sector_count(inode);
2304         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2305         if (status < 0) {
2306                 mlog_errno(status);
2307                 goto bail;
2308         }
2309
2310         status = 0;
2311         if (ret_new_bh) {
2312                 *ret_new_bh = new_bh;
2313                 new_bh = NULL;
2314         }
2315 bail:
2316         brelse(new_bh);
2317
2318         return status;
2319 }
2320
2321 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2322                                      handle_t *handle, struct inode *dir,
2323                                      struct buffer_head *di_bh,
2324                                      struct buffer_head *dirdata_bh,
2325                                      struct ocfs2_alloc_context *meta_ac,
2326                                      int dx_inline, u32 num_entries,
2327                                      struct buffer_head **ret_dx_root_bh)
2328 {
2329         int ret;
2330         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2331         u16 dr_suballoc_bit;
2332         u64 suballoc_loc, dr_blkno;
2333         unsigned int num_bits;
2334         struct buffer_head *dx_root_bh = NULL;
2335         struct ocfs2_dx_root_block *dx_root;
2336         struct ocfs2_dir_block_trailer *trailer =
2337                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2338
2339         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2340                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2341         if (ret) {
2342                 mlog_errno(ret);
2343                 goto out;
2344         }
2345
2346         trace_ocfs2_dx_dir_attach_index(
2347                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2348                                 (unsigned long long)dr_blkno);
2349
2350         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2351         if (dx_root_bh == NULL) {
2352                 ret = -ENOMEM;
2353                 goto out;
2354         }
2355         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2356
2357         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2358                                       OCFS2_JOURNAL_ACCESS_CREATE);
2359         if (ret < 0) {
2360                 mlog_errno(ret);
2361                 goto out;
2362         }
2363
2364         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2365         memset(dx_root, 0, osb->sb->s_blocksize);
2366         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2367         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2368         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2369         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2370         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2371         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2372         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2373         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2374         if (le16_to_cpu(trailer->db_free_rec_len))
2375                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2376         else
2377                 dx_root->dr_free_blk = cpu_to_le64(0);
2378
2379         if (dx_inline) {
2380                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2381                 dx_root->dr_entries.de_count =
2382                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2383         } else {
2384                 dx_root->dr_list.l_count =
2385                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2386         }
2387         ocfs2_journal_dirty(handle, dx_root_bh);
2388
2389         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2390                                       OCFS2_JOURNAL_ACCESS_CREATE);
2391         if (ret) {
2392                 mlog_errno(ret);
2393                 goto out;
2394         }
2395
2396         di->i_dx_root = cpu_to_le64(dr_blkno);
2397
2398         spin_lock(&OCFS2_I(dir)->ip_lock);
2399         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2400         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2401         spin_unlock(&OCFS2_I(dir)->ip_lock);
2402
2403         ocfs2_journal_dirty(handle, di_bh);
2404
2405         *ret_dx_root_bh = dx_root_bh;
2406         dx_root_bh = NULL;
2407
2408 out:
2409         brelse(dx_root_bh);
2410         return ret;
2411 }
2412
2413 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2414                                        handle_t *handle, struct inode *dir,
2415                                        struct buffer_head **dx_leaves,
2416                                        int num_dx_leaves, u64 start_blk)
2417 {
2418         int ret, i;
2419         struct ocfs2_dx_leaf *dx_leaf;
2420         struct buffer_head *bh;
2421
2422         for (i = 0; i < num_dx_leaves; i++) {
2423                 bh = sb_getblk(osb->sb, start_blk + i);
2424                 if (bh == NULL) {
2425                         ret = -ENOMEM;
2426                         goto out;
2427                 }
2428                 dx_leaves[i] = bh;
2429
2430                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2431
2432                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2433                                               OCFS2_JOURNAL_ACCESS_CREATE);
2434                 if (ret < 0) {
2435                         mlog_errno(ret);
2436                         goto out;
2437                 }
2438
2439                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2440
2441                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2442                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2443                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2444                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2445                 dx_leaf->dl_list.de_count =
2446                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2447
2448                 trace_ocfs2_dx_dir_format_cluster(
2449                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2450                                 (unsigned long long)bh->b_blocknr,
2451                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2452
2453                 ocfs2_journal_dirty(handle, bh);
2454         }
2455
2456         ret = 0;
2457 out:
2458         return ret;
2459 }
2460
2461 /*
2462  * Allocates and formats a new cluster for use in an indexed dir
2463  * leaf. This version will not do the extent insert, so that it can be
2464  * used by operations which need careful ordering.
2465  */
2466 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2467                                       u32 cpos, handle_t *handle,
2468                                       struct ocfs2_alloc_context *data_ac,
2469                                       struct buffer_head **dx_leaves,
2470                                       int num_dx_leaves, u64 *ret_phys_blkno)
2471 {
2472         int ret;
2473         u32 phys, num;
2474         u64 phys_blkno;
2475         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2476
2477         /*
2478          * XXX: For create, this should claim cluster for the index
2479          * *before* the unindexed insert so that we have a better
2480          * chance of contiguousness as the directory grows in number
2481          * of entries.
2482          */
2483         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2484         if (ret) {
2485                 mlog_errno(ret);
2486                 goto out;
2487         }
2488
2489         /*
2490          * Format the new cluster first. That way, we're inserting
2491          * valid data.
2492          */
2493         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2494         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2495                                           num_dx_leaves, phys_blkno);
2496         if (ret) {
2497                 mlog_errno(ret);
2498                 goto out;
2499         }
2500
2501         *ret_phys_blkno = phys_blkno;
2502 out:
2503         return ret;
2504 }
2505
2506 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2507                                     struct ocfs2_extent_tree *et,
2508                                     u32 cpos, handle_t *handle,
2509                                     struct ocfs2_alloc_context *data_ac,
2510                                     struct ocfs2_alloc_context *meta_ac,
2511                                     struct buffer_head **dx_leaves,
2512                                     int num_dx_leaves)
2513 {
2514         int ret;
2515         u64 phys_blkno;
2516
2517         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2518                                          num_dx_leaves, &phys_blkno);
2519         if (ret) {
2520                 mlog_errno(ret);
2521                 goto out;
2522         }
2523
2524         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2525                                   meta_ac);
2526         if (ret)
2527                 mlog_errno(ret);
2528 out:
2529         return ret;
2530 }
2531
2532 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2533                                                         int *ret_num_leaves)
2534 {
2535         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2536         struct buffer_head **dx_leaves;
2537
2538         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2539                             GFP_NOFS);
2540         if (dx_leaves && ret_num_leaves)
2541                 *ret_num_leaves = num_dx_leaves;
2542
2543         return dx_leaves;
2544 }
2545
2546 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2547                                  handle_t *handle,
2548                                  struct inode *parent,
2549                                  struct inode *inode,
2550                                  struct buffer_head *di_bh,
2551                                  struct ocfs2_alloc_context *data_ac,
2552                                  struct ocfs2_alloc_context *meta_ac)
2553 {
2554         int ret;
2555         struct buffer_head *leaf_bh = NULL;
2556         struct buffer_head *dx_root_bh = NULL;
2557         struct ocfs2_dx_hinfo hinfo;
2558         struct ocfs2_dx_root_block *dx_root;
2559         struct ocfs2_dx_entry_list *entry_list;
2560
2561         /*
2562          * Our strategy is to create the directory as though it were
2563          * unindexed, then add the index block. This works with very
2564          * little complication since the state of a new directory is a
2565          * very well known quantity.
2566          *
2567          * Essentially, we have two dirents ("." and ".."), in the 1st
2568          * block which need indexing. These are easily inserted into
2569          * the index block.
2570          */
2571
2572         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2573                                     data_ac, &leaf_bh);
2574         if (ret) {
2575                 mlog_errno(ret);
2576                 goto out;
2577         }
2578
2579         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2580                                         meta_ac, 1, 2, &dx_root_bh);
2581         if (ret) {
2582                 mlog_errno(ret);
2583                 goto out;
2584         }
2585         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2586         entry_list = &dx_root->dr_entries;
2587
2588         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2589         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2590         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2591
2592         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2593         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2594
2595 out:
2596         brelse(dx_root_bh);
2597         brelse(leaf_bh);
2598         return ret;
2599 }
2600
2601 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2602                        handle_t *handle,
2603                        struct inode *parent,
2604                        struct inode *inode,
2605                        struct buffer_head *fe_bh,
2606                        struct ocfs2_alloc_context *data_ac,
2607                        struct ocfs2_alloc_context *meta_ac)
2608
2609 {
2610         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2611
2612         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2613                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2614
2615         if (ocfs2_supports_indexed_dirs(osb))
2616                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2617                                              data_ac, meta_ac);
2618
2619         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2620                                      data_ac, NULL);
2621 }
2622
2623 static int ocfs2_dx_dir_index_block(struct inode *dir,
2624                                     handle_t *handle,
2625                                     struct buffer_head **dx_leaves,
2626                                     int num_dx_leaves,
2627                                     u32 *num_dx_entries,
2628                                     struct buffer_head *dirent_bh)
2629 {
2630         int ret = 0, namelen, i;
2631         char *de_buf, *limit;
2632         struct ocfs2_dir_entry *de;
2633         struct buffer_head *dx_leaf_bh;
2634         struct ocfs2_dx_hinfo hinfo;
2635         u64 dirent_blk = dirent_bh->b_blocknr;
2636
2637         de_buf = dirent_bh->b_data;
2638         limit = de_buf + dir->i_sb->s_blocksize;
2639
2640         while (de_buf < limit) {
2641                 de = (struct ocfs2_dir_entry *)de_buf;
2642
2643                 namelen = de->name_len;
2644                 if (!namelen || !de->inode)
2645                         goto inc;
2646
2647                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2648
2649                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2650                 dx_leaf_bh = dx_leaves[i];
2651
2652                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2653                                                  dirent_blk, dx_leaf_bh);
2654                 if (ret) {
2655                         mlog_errno(ret);
2656                         goto out;
2657                 }
2658
2659                 *num_dx_entries = *num_dx_entries + 1;
2660
2661 inc:
2662                 de_buf += le16_to_cpu(de->rec_len);
2663         }
2664
2665 out:
2666         return ret;
2667 }
2668
2669 /*
2670  * XXX: This expects dx_root_bh to already be part of the transaction.
2671  */
2672 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2673                                          struct buffer_head *dx_root_bh,
2674                                          struct buffer_head *dirent_bh)
2675 {
2676         char *de_buf, *limit;
2677         struct ocfs2_dx_root_block *dx_root;
2678         struct ocfs2_dir_entry *de;
2679         struct ocfs2_dx_hinfo hinfo;
2680         u64 dirent_blk = dirent_bh->b_blocknr;
2681
2682         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2683
2684         de_buf = dirent_bh->b_data;
2685         limit = de_buf + dir->i_sb->s_blocksize;
2686
2687         while (de_buf < limit) {
2688                 de = (struct ocfs2_dir_entry *)de_buf;
2689
2690                 if (!de->name_len || !de->inode)
2691                         goto inc;
2692
2693                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2694
2695                 trace_ocfs2_dx_dir_index_root_block(
2696                                 (unsigned long long)dir->i_ino,
2697                                 hinfo.major_hash, hinfo.minor_hash,
2698                                 de->name_len, de->name,
2699                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2700
2701                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2702                                            dirent_blk);
2703
2704                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2705 inc:
2706                 de_buf += le16_to_cpu(de->rec_len);
2707         }
2708 }
2709
2710 /*
2711  * Count the number of inline directory entries in di_bh and compare
2712  * them against the number of entries we can hold in an inline dx root
2713  * block.
2714  */
2715 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2716                                          struct buffer_head *di_bh)
2717 {
2718         int dirent_count = 0;
2719         char *de_buf, *limit;
2720         struct ocfs2_dir_entry *de;
2721         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2722
2723         de_buf = di->id2.i_data.id_data;
2724         limit = de_buf + i_size_read(dir);
2725
2726         while (de_buf < limit) {
2727                 de = (struct ocfs2_dir_entry *)de_buf;
2728
2729                 if (de->name_len && de->inode)
2730                         dirent_count++;
2731
2732                 de_buf += le16_to_cpu(de->rec_len);
2733         }
2734
2735         /* We are careful to leave room for one extra record. */
2736         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2737 }
2738
2739 /*
2740  * Expand rec_len of the rightmost dirent in a directory block so that it
2741  * contains the end of our valid space for dirents. We do this during
2742  * expansion from an inline directory to one with extents. The first dir block
2743  * in that case is taken from the inline data portion of the inode block.
2744  *
2745  * This will also return the largest amount of contiguous space for a dirent
2746  * in the block. That value is *not* necessarily the last dirent, even after
2747  * expansion. The directory indexing code wants this value for free space
2748  * accounting. We do this here since we're already walking the entire dir
2749  * block.
2750  *
2751  * We add the dir trailer if this filesystem wants it.
2752  */
2753 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2754                                              struct inode *dir)
2755 {
2756         struct super_block *sb = dir->i_sb;
2757         struct ocfs2_dir_entry *de;
2758         struct ocfs2_dir_entry *prev_de;
2759         char *de_buf, *limit;
2760         unsigned int new_size = sb->s_blocksize;
2761         unsigned int bytes, this_hole;
2762         unsigned int largest_hole = 0;
2763
2764         if (ocfs2_new_dir_wants_trailer(dir))
2765                 new_size = ocfs2_dir_trailer_blk_off(sb);
2766
2767         bytes = new_size - old_size;
2768
2769         limit = start + old_size;
2770         de_buf = start;
2771         de = (struct ocfs2_dir_entry *)de_buf;
2772         do {
2773                 this_hole = ocfs2_figure_dirent_hole(de);
2774                 if (this_hole > largest_hole)
2775                         largest_hole = this_hole;
2776
2777                 prev_de = de;
2778                 de_buf += le16_to_cpu(de->rec_len);
2779                 de = (struct ocfs2_dir_entry *)de_buf;
2780         } while (de_buf < limit);
2781
2782         le16_add_cpu(&prev_de->rec_len, bytes);
2783
2784         /* We need to double check this after modification of the final
2785          * dirent. */
2786         this_hole = ocfs2_figure_dirent_hole(prev_de);
2787         if (this_hole > largest_hole)
2788                 largest_hole = this_hole;
2789
2790         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2791                 return largest_hole;
2792         return 0;
2793 }
2794
2795 /*
2796  * We allocate enough clusters to fulfill "blocks_wanted", but set
2797  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2798  * rest automatically for us.
2799  *
2800  * *first_block_bh is a pointer to the 1st data block allocated to the
2801  *  directory.
2802  */
2803 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2804                                    unsigned int blocks_wanted,
2805                                    struct ocfs2_dir_lookup_result *lookup,
2806                                    struct buffer_head **first_block_bh)
2807 {
2808         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2809         struct super_block *sb = dir->i_sb;
2810         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2811                 credits = ocfs2_inline_to_extents_credits(sb);
2812         u64 dx_insert_blkno, blkno,
2813                 bytes = blocks_wanted << sb->s_blocksize_bits;
2814         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2815         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2816         struct ocfs2_alloc_context *data_ac = NULL;
2817         struct ocfs2_alloc_context *meta_ac = NULL;
2818         struct buffer_head *dirdata_bh = NULL;
2819         struct buffer_head *dx_root_bh = NULL;
2820         struct buffer_head **dx_leaves = NULL;
2821         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2822         handle_t *handle;
2823         struct ocfs2_extent_tree et;
2824         struct ocfs2_extent_tree dx_et;
2825         int did_quota = 0, bytes_allocated = 0;
2826
2827         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2828
2829         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2830         dx_alloc = 0;
2831
2832         down_write(&oi->ip_alloc_sem);
2833
2834         if (ocfs2_supports_indexed_dirs(osb)) {
2835                 credits += ocfs2_add_dir_index_credits(sb);
2836
2837                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2838                 if (!dx_inline) {
2839                         /* Add one more cluster for an index leaf */
2840                         dx_alloc++;
2841                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2842                                                                 &num_dx_leaves);
2843                         if (!dx_leaves) {
2844                                 ret = -ENOMEM;
2845                                 mlog_errno(ret);
2846                                 goto out;
2847                         }
2848                 }
2849
2850                 /* This gets us the dx_root */
2851                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2852                 if (ret) {
2853                         mlog_errno(ret);
2854                         goto out;
2855                 }
2856         }
2857
2858         /*
2859          * We should never need more than 2 clusters for the unindexed
2860          * tree - maximum dirent size is far less than one block. In
2861          * fact, the only time we'd need more than one cluster is if
2862          * blocksize == clustersize and the dirent won't fit in the
2863          * extra space that the expansion to a single block gives. As
2864          * of today, that only happens on 4k/4k file systems.
2865          */
2866         BUG_ON(alloc > 2);
2867
2868         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2869         if (ret) {
2870                 mlog_errno(ret);
2871                 goto out;
2872         }
2873
2874         /*
2875          * Prepare for worst case allocation scenario of two separate
2876          * extents in the unindexed tree.
2877          */
2878         if (alloc == 2)
2879                 credits += OCFS2_SUBALLOC_ALLOC;
2880
2881         handle = ocfs2_start_trans(osb, credits);
2882         if (IS_ERR(handle)) {
2883                 ret = PTR_ERR(handle);
2884                 mlog_errno(ret);
2885                 goto out;
2886         }
2887
2888         ret = dquot_alloc_space_nodirty(dir,
2889                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2890         if (ret)
2891                 goto out_commit;
2892         did_quota = 1;
2893
2894         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2895                 /*
2896                  * Allocate our index cluster first, to maximize the
2897                  * possibility that unindexed leaves grow
2898                  * contiguously.
2899                  */
2900                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2901                                                  dx_leaves, num_dx_leaves,
2902                                                  &dx_insert_blkno);
2903                 if (ret) {
2904                         mlog_errno(ret);
2905                         goto out_commit;
2906                 }
2907                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2908         }
2909
2910         /*
2911          * Try to claim as many clusters as the bitmap can give though
2912          * if we only get one now, that's enough to continue. The rest
2913          * will be claimed after the conversion to extents.
2914          */
2915         if (ocfs2_dir_resv_allowed(osb))
2916                 data_ac->ac_resv = &oi->ip_la_data_resv;
2917         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2918         if (ret) {
2919                 mlog_errno(ret);
2920                 goto out_commit;
2921         }
2922         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2923
2924         /*
2925          * Operations are carefully ordered so that we set up the new
2926          * data block first. The conversion from inline data to
2927          * extents follows.
2928          */
2929         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2930         dirdata_bh = sb_getblk(sb, blkno);
2931         if (!dirdata_bh) {
2932                 ret = -ENOMEM;
2933                 mlog_errno(ret);
2934                 goto out_commit;
2935         }
2936
2937         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2938
2939         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2940                                       OCFS2_JOURNAL_ACCESS_CREATE);
2941         if (ret) {
2942                 mlog_errno(ret);
2943                 goto out_commit;
2944         }
2945
2946         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2947         memset(dirdata_bh->b_data + i_size_read(dir), 0,
2948                sb->s_blocksize - i_size_read(dir));
2949         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2950         if (ocfs2_new_dir_wants_trailer(dir)) {
2951                 /*
2952                  * Prepare the dir trailer up front. It will otherwise look
2953                  * like a valid dirent. Even if inserting the index fails
2954                  * (unlikely), then all we'll have done is given first dir
2955                  * block a small amount of fragmentation.
2956                  */
2957                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2958         }
2959
2960         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2961         ocfs2_journal_dirty(handle, dirdata_bh);
2962
2963         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2964                 /*
2965                  * Dx dirs with an external cluster need to do this up
2966                  * front. Inline dx root's get handled later, after
2967                  * we've allocated our root block. We get passed back
2968                  * a total number of items so that dr_num_entries can
2969                  * be correctly set once the dx_root has been
2970                  * allocated.
2971                  */
2972                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2973                                                num_dx_leaves, &num_dx_entries,
2974                                                dirdata_bh);
2975                 if (ret) {
2976                         mlog_errno(ret);
2977                         goto out_commit;
2978                 }
2979         }
2980
2981         /*
2982          * Set extent, i_size, etc on the directory. After this, the
2983          * inode should contain the same exact dirents as before and
2984          * be fully accessible from system calls.
2985          *
2986          * We let the later dirent insert modify c/mtime - to the user
2987          * the data hasn't changed.
2988          */
2989         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2990                                       OCFS2_JOURNAL_ACCESS_CREATE);
2991         if (ret) {
2992                 mlog_errno(ret);
2993                 goto out_commit;
2994         }
2995
2996         spin_lock(&oi->ip_lock);
2997         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
2998         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
2999         spin_unlock(&oi->ip_lock);
3000
3001         ocfs2_dinode_new_extent_list(dir, di);
3002
3003         i_size_write(dir, sb->s_blocksize);
3004         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3005
3006         di->i_size = cpu_to_le64(sb->s_blocksize);
3007         di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3008         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3009         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3010
3011         /*
3012          * This should never fail as our extent list is empty and all
3013          * related blocks have been journaled already.
3014          */
3015         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
3016                                   0, NULL);
3017         if (ret) {
3018                 mlog_errno(ret);
3019                 goto out_commit;
3020         }
3021
3022         /*
3023          * Set i_blocks after the extent insert for the most up to
3024          * date ip_clusters value.
3025          */
3026         dir->i_blocks = ocfs2_inode_sector_count(dir);
3027
3028         ocfs2_journal_dirty(handle, di_bh);
3029
3030         if (ocfs2_supports_indexed_dirs(osb)) {
3031                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3032                                                 dirdata_bh, meta_ac, dx_inline,
3033                                                 num_dx_entries, &dx_root_bh);
3034                 if (ret) {
3035                         mlog_errno(ret);
3036                         goto out_commit;
3037                 }
3038
3039                 if (dx_inline) {
3040                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3041                                                       dirdata_bh);
3042                 } else {
3043                         ocfs2_init_dx_root_extent_tree(&dx_et,
3044                                                        INODE_CACHE(dir),
3045                                                        dx_root_bh);
3046                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3047                                                   dx_insert_blkno, 1, 0, NULL);
3048                         if (ret)
3049                                 mlog_errno(ret);
3050                 }
3051         }
3052
3053         /*
3054          * We asked for two clusters, but only got one in the 1st
3055          * pass. Claim the 2nd cluster as a separate extent.
3056          */
3057         if (alloc > len) {
3058                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3059                                            &len);
3060                 if (ret) {
3061                         mlog_errno(ret);
3062                         goto out_commit;
3063                 }
3064                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3065
3066                 ret = ocfs2_insert_extent(handle, &et, 1,
3067                                           blkno, len, 0, NULL);
3068                 if (ret) {
3069                         mlog_errno(ret);
3070                         goto out_commit;
3071                 }
3072                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3073         }
3074
3075         *first_block_bh = dirdata_bh;
3076         dirdata_bh = NULL;
3077         if (ocfs2_supports_indexed_dirs(osb)) {
3078                 unsigned int off;
3079
3080                 if (!dx_inline) {
3081                         /*
3082                          * We need to return the correct block within the
3083                          * cluster which should hold our entry.
3084                          */
3085                         off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3086                                                     &lookup->dl_hinfo);
3087                         get_bh(dx_leaves[off]);
3088                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3089                 }
3090                 lookup->dl_dx_root_bh = dx_root_bh;
3091                 dx_root_bh = NULL;
3092         }
3093
3094 out_commit:
3095         if (ret < 0 && did_quota)
3096                 dquot_free_space_nodirty(dir, bytes_allocated);
3097
3098         ocfs2_commit_trans(osb, handle);
3099
3100 out:
3101         up_write(&oi->ip_alloc_sem);
3102         if (data_ac)
3103                 ocfs2_free_alloc_context(data_ac);
3104         if (meta_ac)
3105                 ocfs2_free_alloc_context(meta_ac);
3106
3107         if (dx_leaves) {
3108                 for (i = 0; i < num_dx_leaves; i++)
3109                         brelse(dx_leaves[i]);
3110                 kfree(dx_leaves);
3111         }
3112
3113         brelse(dirdata_bh);
3114         brelse(dx_root_bh);
3115
3116         return ret;
3117 }
3118
3119 /* returns a bh of the 1st new block in the allocation. */
3120 static int ocfs2_do_extend_dir(struct super_block *sb,
3121                                handle_t *handle,
3122                                struct inode *dir,
3123                                struct buffer_head *parent_fe_bh,
3124                                struct ocfs2_alloc_context *data_ac,
3125                                struct ocfs2_alloc_context *meta_ac,
3126                                struct buffer_head **new_bh)
3127 {
3128         int status;
3129         int extend, did_quota = 0;
3130         u64 p_blkno, v_blkno;
3131
3132         spin_lock(&OCFS2_I(dir)->ip_lock);
3133         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3134         spin_unlock(&OCFS2_I(dir)->ip_lock);
3135
3136         if (extend) {
3137                 u32 offset = OCFS2_I(dir)->ip_clusters;
3138
3139                 status = dquot_alloc_space_nodirty(dir,
3140                                         ocfs2_clusters_to_bytes(sb, 1));
3141                 if (status)
3142                         goto bail;
3143                 did_quota = 1;
3144
3145                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3146                                               1, 0, parent_fe_bh, handle,
3147                                               data_ac, meta_ac, NULL);
3148                 BUG_ON(status == -EAGAIN);
3149                 if (status < 0) {
3150                         mlog_errno(status);
3151                         goto bail;
3152                 }
3153         }
3154
3155         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3156         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3157         if (status < 0) {
3158                 mlog_errno(status);
3159                 goto bail;
3160         }
3161
3162         *new_bh = sb_getblk(sb, p_blkno);
3163         if (!*new_bh) {
3164                 status = -ENOMEM;
3165                 mlog_errno(status);
3166                 goto bail;
3167         }
3168         status = 0;
3169 bail:
3170         if (did_quota && status < 0)
3171                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3172         return status;
3173 }
3174
3175 /*
3176  * Assumes you already have a cluster lock on the directory.
3177  *
3178  * 'blocks_wanted' is only used if we have an inline directory which
3179  * is to be turned into an extent based one. The size of the dirent to
3180  * insert might be larger than the space gained by growing to just one
3181  * block, so we may have to grow the inode by two blocks in that case.
3182  *
3183  * If the directory is already indexed, dx_root_bh must be provided.
3184  */
3185 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3186                             struct inode *dir,
3187                             struct buffer_head *parent_fe_bh,
3188                             unsigned int blocks_wanted,
3189                             struct ocfs2_dir_lookup_result *lookup,
3190                             struct buffer_head **new_de_bh)
3191 {
3192         int status = 0;
3193         int credits, num_free_extents, drop_alloc_sem = 0;
3194         loff_t dir_i_size;
3195         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3196         struct ocfs2_extent_list *el = &fe->id2.i_list;
3197         struct ocfs2_alloc_context *data_ac = NULL;
3198         struct ocfs2_alloc_context *meta_ac = NULL;
3199         handle_t *handle = NULL;
3200         struct buffer_head *new_bh = NULL;
3201         struct ocfs2_dir_entry * de;
3202         struct super_block *sb = osb->sb;
3203         struct ocfs2_extent_tree et;
3204         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3205
3206         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3207                 /*
3208                  * This would be a code error as an inline directory should
3209                  * never have an index root.
3210                  */
3211                 BUG_ON(dx_root_bh);
3212
3213                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3214                                                  blocks_wanted, lookup,
3215                                                  &new_bh);
3216                 if (status) {
3217                         mlog_errno(status);
3218                         goto bail;
3219                 }
3220
3221                 /* Expansion from inline to an indexed directory will
3222                  * have given us this. */
3223                 dx_root_bh = lookup->dl_dx_root_bh;
3224
3225                 if (blocks_wanted == 1) {
3226                         /*
3227                          * If the new dirent will fit inside the space
3228                          * created by pushing out to one block, then
3229                          * we can complete the operation
3230                          * here. Otherwise we have to expand i_size
3231                          * and format the 2nd block below.
3232                          */
3233                         BUG_ON(new_bh == NULL);
3234                         goto bail_bh;
3235                 }
3236
3237                 /*
3238                  * Get rid of 'new_bh' - we want to format the 2nd
3239                  * data block and return that instead.
3240                  */
3241                 brelse(new_bh);
3242                 new_bh = NULL;
3243
3244                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3245                 drop_alloc_sem = 1;
3246                 dir_i_size = i_size_read(dir);
3247                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3248                 goto do_extend;
3249         }
3250
3251         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3252         drop_alloc_sem = 1;
3253         dir_i_size = i_size_read(dir);
3254         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3255                                dir_i_size);
3256
3257         /* dir->i_size is always block aligned. */
3258         spin_lock(&OCFS2_I(dir)->ip_lock);
3259         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3260                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3261                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3262                                               parent_fe_bh);
3263                 num_free_extents = ocfs2_num_free_extents(osb, &et);
3264                 if (num_free_extents < 0) {
3265                         status = num_free_extents;
3266                         mlog_errno(status);
3267                         goto bail;
3268                 }
3269
3270                 if (!num_free_extents) {
3271                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3272                         if (status < 0) {
3273                                 if (status != -ENOSPC)
3274                                         mlog_errno(status);
3275                                 goto bail;
3276                         }
3277                 }
3278
3279                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3280                 if (status < 0) {
3281                         if (status != -ENOSPC)
3282                                 mlog_errno(status);
3283                         goto bail;
3284                 }
3285
3286                 if (ocfs2_dir_resv_allowed(osb))
3287                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3288
3289                 credits = ocfs2_calc_extend_credits(sb, el);
3290         } else {
3291                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3292                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3293         }
3294
3295 do_extend:
3296         if (ocfs2_dir_indexed(dir))
3297                 credits++; /* For attaching the new dirent block to the
3298                             * dx_root */
3299
3300         handle = ocfs2_start_trans(osb, credits);
3301         if (IS_ERR(handle)) {
3302                 status = PTR_ERR(handle);
3303                 handle = NULL;
3304                 mlog_errno(status);
3305                 goto bail;
3306         }
3307
3308         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3309                                      data_ac, meta_ac, &new_bh);
3310         if (status < 0) {
3311                 mlog_errno(status);
3312                 goto bail;
3313         }
3314
3315         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3316
3317         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3318                                          OCFS2_JOURNAL_ACCESS_CREATE);
3319         if (status < 0) {
3320                 mlog_errno(status);
3321                 goto bail;
3322         }
3323         memset(new_bh->b_data, 0, sb->s_blocksize);
3324
3325         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3326         de->inode = 0;
3327         if (ocfs2_supports_dir_trailer(dir)) {
3328                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3329
3330                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3331
3332                 if (ocfs2_dir_indexed(dir)) {
3333                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3334                                                            dx_root_bh, new_bh);
3335                         if (status) {
3336                                 mlog_errno(status);
3337                                 goto bail;
3338                         }
3339                 }
3340         } else {
3341                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3342         }
3343         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3344         ocfs2_journal_dirty(handle, new_bh);
3345
3346         dir_i_size += dir->i_sb->s_blocksize;
3347         i_size_write(dir, dir_i_size);
3348         dir->i_blocks = ocfs2_inode_sector_count(dir);
3349         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3350         if (status < 0) {
3351                 mlog_errno(status);
3352                 goto bail;
3353         }
3354
3355 bail_bh:
3356         *new_de_bh = new_bh;
3357         get_bh(*new_de_bh);
3358 bail:
3359         if (handle)
3360                 ocfs2_commit_trans(osb, handle);
3361         if (drop_alloc_sem)
3362                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3363
3364         if (data_ac)
3365                 ocfs2_free_alloc_context(data_ac);
3366         if (meta_ac)
3367                 ocfs2_free_alloc_context(meta_ac);
3368
3369         brelse(new_bh);
3370
3371         return status;
3372 }
3373
3374 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3375                                    const char *name, int namelen,
3376                                    struct buffer_head **ret_de_bh,
3377                                    unsigned int *blocks_wanted)
3378 {
3379         int ret;
3380         struct super_block *sb = dir->i_sb;
3381         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3382         struct ocfs2_dir_entry *de, *last_de = NULL;
3383         char *de_buf, *limit;
3384         unsigned long offset = 0;
3385         unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3386
3387         /*
3388          * This calculates how many free bytes we'd have in block zero, should
3389          * this function force expansion to an extent tree.
3390          */
3391         if (ocfs2_new_dir_wants_trailer(dir))
3392                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3393         else
3394                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3395
3396         de_buf = di->id2.i_data.id_data;
3397         limit = de_buf + i_size_read(dir);
3398         rec_len = OCFS2_DIR_REC_LEN(namelen);
3399
3400         while (de_buf < limit) {
3401                 de = (struct ocfs2_dir_entry *)de_buf;
3402
3403                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3404                         ret = -ENOENT;
3405                         goto out;
3406                 }
3407                 if (ocfs2_match(namelen, name, de)) {
3408                         ret = -EEXIST;
3409                         goto out;
3410                 }
3411                 /*
3412                  * No need to check for a trailing dirent record here as
3413                  * they're not used for inline dirs.
3414                  */
3415
3416                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3417                         /* Ok, we found a spot. Return this bh and let
3418                          * the caller actually fill it in. */
3419                         *ret_de_bh = di_bh;
3420                         get_bh(*ret_de_bh);
3421                         ret = 0;
3422                         goto out;
3423                 }
3424
3425                 last_de = de;
3426                 de_buf += le16_to_cpu(de->rec_len);
3427                 offset += le16_to_cpu(de->rec_len);
3428         }
3429
3430         /*
3431          * We're going to require expansion of the directory - figure
3432          * out how many blocks we'll need so that a place for the
3433          * dirent can be found.
3434          */
3435         *blocks_wanted = 1;
3436         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3437         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3438                 *blocks_wanted = 2;
3439
3440         ret = -ENOSPC;
3441 out:
3442         return ret;
3443 }
3444
3445 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3446                                    int namelen, struct buffer_head **ret_de_bh)
3447 {
3448         unsigned long offset;
3449         struct buffer_head *bh = NULL;
3450         unsigned short rec_len;
3451         struct ocfs2_dir_entry *de;
3452         struct super_block *sb = dir->i_sb;
3453         int status;
3454         int blocksize = dir->i_sb->s_blocksize;
3455
3456         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3457         if (status) {
3458                 mlog_errno(status);
3459                 goto bail;
3460         }
3461
3462         rec_len = OCFS2_DIR_REC_LEN(namelen);
3463         offset = 0;
3464         de = (struct ocfs2_dir_entry *) bh->b_data;
3465         while (1) {
3466                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3467                         brelse(bh);
3468                         bh = NULL;
3469
3470                         if (i_size_read(dir) <= offset) {
3471                                 /*
3472                                  * Caller will have to expand this
3473                                  * directory.
3474                                  */
3475                                 status = -ENOSPC;
3476                                 goto bail;
3477                         }
3478                         status = ocfs2_read_dir_block(dir,
3479                                              offset >> sb->s_blocksize_bits,
3480                                              &bh, 0);
3481                         if (status) {
3482                                 mlog_errno(status);
3483                                 goto bail;
3484                         }
3485                         /* move to next block */
3486                         de = (struct ocfs2_dir_entry *) bh->b_data;
3487                 }
3488                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3489                         status = -ENOENT;
3490                         goto bail;
3491                 }
3492                 if (ocfs2_match(namelen, name, de)) {
3493                         status = -EEXIST;
3494                         goto bail;
3495                 }
3496
3497                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3498                                            blocksize))
3499                         goto next;
3500
3501                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3502                         /* Ok, we found a spot. Return this bh and let
3503                          * the caller actually fill it in. */
3504                         *ret_de_bh = bh;
3505                         get_bh(*ret_de_bh);
3506                         status = 0;
3507                         goto bail;
3508                 }
3509 next:
3510                 offset += le16_to_cpu(de->rec_len);
3511                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3512         }
3513
3514         status = 0;
3515 bail:
3516         brelse(bh);
3517         if (status)
3518                 mlog_errno(status);
3519
3520         return status;
3521 }
3522
3523 static int dx_leaf_sort_cmp(const void *a, const void *b)
3524 {
3525         const struct ocfs2_dx_entry *entry1 = a;
3526         const struct ocfs2_dx_entry *entry2 = b;
3527         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3528         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3529         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3530         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3531
3532         if (major_hash1 > major_hash2)
3533                 return 1;
3534         if (major_hash1 < major_hash2)
3535                 return -1;
3536
3537         /*
3538          * It is not strictly necessary to sort by minor
3539          */
3540         if (minor_hash1 > minor_hash2)
3541                 return 1;
3542         if (minor_hash1 < minor_hash2)
3543                 return -1;
3544         return 0;
3545 }
3546
3547 static void dx_leaf_sort_swap(void *a, void *b, int size)
3548 {
3549         struct ocfs2_dx_entry *entry1 = a;
3550         struct ocfs2_dx_entry *entry2 = b;
3551         struct ocfs2_dx_entry tmp;
3552
3553         BUG_ON(size != sizeof(*entry1));
3554
3555         tmp = *entry1;
3556         *entry1 = *entry2;
3557         *entry2 = tmp;
3558 }
3559
3560 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3561 {
3562         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3563         int i, num = le16_to_cpu(dl_list->de_num_used);
3564
3565         for (i = 0; i < (num - 1); i++) {
3566                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3567                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3568                         return 0;
3569         }
3570
3571         return 1;
3572 }
3573
3574 /*
3575  * Find the optimal value to split this leaf on. This expects the leaf
3576  * entries to be in sorted order.
3577  *
3578  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3579  * the hash we want to insert.
3580  *
3581  * This function is only concerned with the major hash - that which
3582  * determines which cluster an item belongs to.
3583  */
3584 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3585                                         u32 leaf_cpos, u32 insert_hash,
3586                                         u32 *split_hash)
3587 {
3588         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3589         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3590         int allsame;
3591
3592         /*
3593          * There's a couple rare, but nasty corner cases we have to
3594          * check for here. All of them involve a leaf where all value
3595          * have the same hash, which is what we look for first.
3596          *
3597          * Most of the time, all of the above is false, and we simply
3598          * pick the median value for a split.
3599          */
3600         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3601         if (allsame) {
3602                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3603
3604                 if (val == insert_hash) {
3605                         /*
3606                          * No matter where we would choose to split,
3607                          * the new entry would want to occupy the same
3608                          * block as these. Since there's no space left
3609                          * in their existing block, we know there
3610                          * won't be space after the split.
3611                          */
3612                         return -ENOSPC;
3613                 }
3614
3615                 if (val == leaf_cpos) {
3616                         /*
3617                          * Because val is the same as leaf_cpos (which
3618                          * is the smallest value this leaf can have),
3619                          * yet is not equal to insert_hash, then we
3620                          * know that insert_hash *must* be larger than
3621                          * val (and leaf_cpos). At least cpos+1 in value.
3622                          *
3623                          * We also know then, that there cannot be an
3624                          * adjacent extent (otherwise we'd be looking
3625                          * at it). Choosing this value gives us a
3626                          * chance to get some contiguousness.
3627                          */
3628                         *split_hash = leaf_cpos + 1;
3629                         return 0;
3630                 }
3631
3632                 if (val > insert_hash) {
3633                         /*
3634                          * val can not be the same as insert hash, and
3635                          * also must be larger than leaf_cpos. Also,
3636                          * we know that there can't be a leaf between
3637                          * cpos and val, otherwise the entries with
3638                          * hash 'val' would be there.
3639                          */
3640                         *split_hash = val;
3641                         return 0;
3642                 }
3643
3644                 *split_hash = insert_hash;
3645                 return 0;
3646         }
3647
3648         /*
3649          * Since the records are sorted and the checks above
3650          * guaranteed that not all records in this block are the same,
3651          * we simple travel forward, from the median, and pick the 1st
3652          * record whose value is larger than leaf_cpos.
3653          */
3654         for (i = (num_used / 2); i < num_used; i++)
3655                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3656                     leaf_cpos)
3657                         break;
3658
3659         BUG_ON(i == num_used); /* Should be impossible */
3660         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3661         return 0;
3662 }
3663
3664 /*
3665  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3666  * larger than split_hash into new_dx_leaves. We use a temporary
3667  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3668  *
3669  * Since the block offset inside a leaf (cluster) is a constant mask
3670  * of minor_hash, we can optimize - an item at block offset X within
3671  * the original cluster, will be at offset X within the new cluster.
3672  */
3673 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3674                                        handle_t *handle,
3675                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3676                                        struct buffer_head **orig_dx_leaves,
3677                                        struct buffer_head **new_dx_leaves,
3678                                        int num_dx_leaves)
3679 {
3680         int i, j, num_used;
3681         u32 major_hash;
3682         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3683         struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3684         struct ocfs2_dx_entry *dx_entry;
3685
3686         tmp_list = &tmp_dx_leaf->dl_list;
3687
3688         for (i = 0; i < num_dx_leaves; i++) {
3689                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3690                 orig_list = &orig_dx_leaf->dl_list;
3691                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3692                 new_list = &new_dx_leaf->dl_list;
3693
3694                 num_used = le16_to_cpu(orig_list->de_num_used);
3695
3696                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3697                 tmp_list->de_num_used = cpu_to_le16(0);
3698                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3699
3700                 for (j = 0; j < num_used; j++) {
3701                         dx_entry = &orig_list->de_entries[j];
3702                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3703                         if (major_hash >= split_hash)
3704                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3705                                                               dx_entry);
3706                         else
3707                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3708                                                               dx_entry);
3709                 }
3710                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3711
3712                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3713                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3714         }
3715 }
3716
3717 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3718                                           struct ocfs2_dx_root_block *dx_root)
3719 {
3720         int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3721
3722         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3723         credits += ocfs2_quota_trans_credits(osb->sb);
3724         return credits;
3725 }
3726
3727 /*
3728  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3729  * half our entries into.
3730  */
3731 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3732                                   struct buffer_head *dx_root_bh,
3733                                   struct buffer_head *dx_leaf_bh,
3734                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3735                                   u64 leaf_blkno)
3736 {
3737         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3738         int credits, ret, i, num_used, did_quota = 0;
3739         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3740         u64 orig_leaves_start;
3741         int num_dx_leaves;
3742         struct buffer_head **orig_dx_leaves = NULL;
3743         struct buffer_head **new_dx_leaves = NULL;
3744         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3745         struct ocfs2_extent_tree et;
3746         handle_t *handle = NULL;
3747         struct ocfs2_dx_root_block *dx_root;
3748         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3749
3750         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3751                                      (unsigned long long)leaf_blkno,
3752                                      insert_hash);
3753
3754         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3755
3756         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3757         /*
3758          * XXX: This is a rather large limit. We should use a more
3759          * realistic value.
3760          */
3761         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3762                 return -ENOSPC;
3763
3764         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3765         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3766                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3767                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3768                      (unsigned long long)leaf_blkno, num_used);
3769                 ret = -EIO;
3770                 goto out;
3771         }
3772
3773         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3774         if (!orig_dx_leaves) {
3775                 ret = -ENOMEM;
3776                 mlog_errno(ret);
3777                 goto out;
3778         }
3779
3780         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3781         if (!new_dx_leaves) {
3782                 ret = -ENOMEM;
3783                 mlog_errno(ret);
3784                 goto out;
3785         }
3786
3787         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3788         if (ret) {
3789                 if (ret != -ENOSPC)
3790                         mlog_errno(ret);
3791                 goto out;
3792         }
3793
3794         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3795         handle = ocfs2_start_trans(osb, credits);
3796         if (IS_ERR(handle)) {
3797                 ret = PTR_ERR(handle);
3798                 handle = NULL;
3799                 mlog_errno(ret);
3800                 goto out;
3801         }
3802
3803         ret = dquot_alloc_space_nodirty(dir,
3804                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3805         if (ret)
3806                 goto out_commit;
3807         did_quota = 1;
3808
3809         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3810                                       OCFS2_JOURNAL_ACCESS_WRITE);
3811         if (ret) {
3812                 mlog_errno(ret);
3813                 goto out_commit;
3814         }
3815
3816         /*
3817          * This block is changing anyway, so we can sort it in place.
3818          */
3819         sort(dx_leaf->dl_list.de_entries, num_used,
3820              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3821              dx_leaf_sort_swap);
3822
3823         ocfs2_journal_dirty(handle, dx_leaf_bh);
3824
3825         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3826                                            &split_hash);
3827         if (ret) {
3828                 mlog_errno(ret);
3829                 goto  out_commit;
3830         }
3831
3832         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3833
3834         /*
3835          * We have to carefully order operations here. There are items
3836          * which want to be in the new cluster before insert, but in
3837          * order to put those items in the new cluster, we alter the
3838          * old cluster. A failure to insert gets nasty.
3839          *
3840          * So, start by reserving writes to the old
3841          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3842          * the new cluster for us, before inserting it. The insert
3843          * won't happen if there's an error before that. Once the
3844          * insert is done then, we can transfer from one leaf into the
3845          * other without fear of hitting any error.
3846          */
3847
3848         /*
3849          * The leaf transfer wants some scratch space so that we don't
3850          * wind up doing a bunch of expensive memmove().
3851          */
3852         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3853         if (!tmp_dx_leaf) {
3854                 ret = -ENOMEM;
3855                 mlog_errno(ret);
3856                 goto out_commit;
3857         }
3858
3859         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3860         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3861                                    orig_dx_leaves);
3862         if (ret) {
3863                 mlog_errno(ret);
3864                 goto out_commit;
3865         }
3866
3867         cpos = split_hash;
3868         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3869                                        data_ac, meta_ac, new_dx_leaves,
3870                                        num_dx_leaves);
3871         if (ret) {
3872                 mlog_errno(ret);
3873                 goto out_commit;
3874         }
3875
3876         for (i = 0; i < num_dx_leaves; i++) {
3877                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3878                                               orig_dx_leaves[i],
3879                                               OCFS2_JOURNAL_ACCESS_WRITE);
3880                 if (ret) {
3881                         mlog_errno(ret);
3882                         goto out_commit;
3883                 }
3884
3885                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3886                                               new_dx_leaves[i],
3887                                               OCFS2_JOURNAL_ACCESS_WRITE);
3888                 if (ret) {
3889                         mlog_errno(ret);
3890                         goto out_commit;
3891                 }
3892         }
3893
3894         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3895                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3896
3897 out_commit:
3898         if (ret < 0 && did_quota)
3899                 dquot_free_space_nodirty(dir,
3900                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3901
3902         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3903         ocfs2_commit_trans(osb, handle);
3904
3905 out:
3906         if (orig_dx_leaves || new_dx_leaves) {
3907                 for (i = 0; i < num_dx_leaves; i++) {
3908                         if (orig_dx_leaves)
3909                                 brelse(orig_dx_leaves[i]);
3910                         if (new_dx_leaves)
3911                                 brelse(new_dx_leaves[i]);
3912                 }
3913                 kfree(orig_dx_leaves);
3914                 kfree(new_dx_leaves);
3915         }
3916
3917         if (meta_ac)
3918                 ocfs2_free_alloc_context(meta_ac);
3919         if (data_ac)
3920                 ocfs2_free_alloc_context(data_ac);
3921
3922         kfree(tmp_dx_leaf);
3923         return ret;
3924 }
3925
3926 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3927                                    struct buffer_head *di_bh,
3928                                    struct buffer_head *dx_root_bh,
3929                                    const char *name, int namelen,
3930                                    struct ocfs2_dir_lookup_result *lookup)
3931 {
3932         int ret, rebalanced = 0;
3933         struct ocfs2_dx_root_block *dx_root;
3934         struct buffer_head *dx_leaf_bh = NULL;
3935         struct ocfs2_dx_leaf *dx_leaf;
3936         u64 blkno;
3937         u32 leaf_cpos;
3938
3939         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3940
3941 restart_search:
3942         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3943                                   &leaf_cpos, &blkno);
3944         if (ret) {
3945                 mlog_errno(ret);
3946                 goto out;
3947         }
3948
3949         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3950         if (ret) {
3951                 mlog_errno(ret);
3952                 goto out;
3953         }
3954
3955         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3956
3957         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3958             le16_to_cpu(dx_leaf->dl_list.de_count)) {
3959                 if (rebalanced) {
3960                         /*
3961                          * Rebalancing should have provided us with
3962                          * space in an appropriate leaf.
3963                          *
3964                          * XXX: Is this an abnormal condition then?
3965                          * Should we print a message here?
3966                          */
3967                         ret = -ENOSPC;
3968                         goto out;
3969                 }
3970
3971                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3972                                              &lookup->dl_hinfo, leaf_cpos,
3973                                              blkno);
3974                 if (ret) {
3975                         if (ret != -ENOSPC)
3976                                 mlog_errno(ret);
3977                         goto out;
3978                 }
3979
3980                 /*
3981                  * Restart the lookup. The rebalance might have
3982                  * changed which block our item fits into. Mark our
3983                  * progress, so we only execute this once.
3984                  */
3985                 brelse(dx_leaf_bh);
3986                 dx_leaf_bh = NULL;
3987                 rebalanced = 1;
3988                 goto restart_search;
3989         }
3990
3991         lookup->dl_dx_leaf_bh = dx_leaf_bh;
3992         dx_leaf_bh = NULL;
3993
3994 out:
3995         brelse(dx_leaf_bh);
3996         return ret;
3997 }
3998
3999 static int ocfs2_search_dx_free_list(struct inode *dir,
4000                                      struct buffer_head *dx_root_bh,
4001                                      int namelen,
4002                                      struct ocfs2_dir_lookup_result *lookup)
4003 {
4004         int ret = -ENOSPC;
4005         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4006         struct ocfs2_dir_block_trailer *db;
4007         u64 next_block;
4008         int rec_len = OCFS2_DIR_REC_LEN(namelen);
4009         struct ocfs2_dx_root_block *dx_root;
4010
4011         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4012         next_block = le64_to_cpu(dx_root->dr_free_blk);
4013
4014         while (next_block) {
4015                 brelse(prev_leaf_bh);
4016                 prev_leaf_bh = leaf_bh;
4017                 leaf_bh = NULL;
4018
4019                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4020                 if (ret) {
4021                         mlog_errno(ret);
4022                         goto out;
4023                 }
4024
4025                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4026                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4027                         lookup->dl_leaf_bh = leaf_bh;
4028                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
4029                         leaf_bh = NULL;
4030                         prev_leaf_bh = NULL;
4031                         break;
4032                 }
4033
4034                 next_block = le64_to_cpu(db->db_free_next);
4035         }
4036
4037         if (!next_block)
4038                 ret = -ENOSPC;
4039
4040 out:
4041
4042         brelse(leaf_bh);
4043         brelse(prev_leaf_bh);
4044         return ret;
4045 }
4046
4047 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4048                                        struct buffer_head *dx_root_bh)
4049 {
4050         int ret, num_dx_leaves, i, j, did_quota = 0;
4051         struct buffer_head **dx_leaves = NULL;
4052         struct ocfs2_extent_tree et;
4053         u64 insert_blkno;
4054         struct ocfs2_alloc_context *data_ac = NULL;
4055         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4056         handle_t *handle = NULL;
4057         struct ocfs2_dx_root_block *dx_root;
4058         struct ocfs2_dx_entry_list *entry_list;
4059         struct ocfs2_dx_entry *dx_entry;
4060         struct ocfs2_dx_leaf *target_leaf;
4061
4062         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4063         if (ret) {
4064                 mlog_errno(ret);
4065                 goto out;
4066         }
4067
4068         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4069         if (!dx_leaves) {
4070                 ret = -ENOMEM;
4071                 mlog_errno(ret);
4072                 goto out;
4073         }
4074
4075         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4076         if (IS_ERR(handle)) {
4077                 ret = PTR_ERR(handle);
4078                 mlog_errno(ret);
4079                 goto out;
4080         }
4081
4082         ret = dquot_alloc_space_nodirty(dir,
4083                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4084         if (ret)
4085                 goto out_commit;
4086         did_quota = 1;
4087
4088         /*
4089          * We do this up front, before the allocation, so that a
4090          * failure to add the dx_root_bh to the journal won't result
4091          * us losing clusters.
4092          */
4093         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4094                                       OCFS2_JOURNAL_ACCESS_WRITE);
4095         if (ret) {
4096                 mlog_errno(ret);
4097                 goto out_commit;
4098         }
4099
4100         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4101                                          num_dx_leaves, &insert_blkno);
4102         if (ret) {
4103                 mlog_errno(ret);
4104                 goto out_commit;
4105         }
4106
4107         /*
4108          * Transfer the entries from our dx_root into the appropriate
4109          * block
4110          */
4111         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4112         entry_list = &dx_root->dr_entries;
4113
4114         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4115                 dx_entry = &entry_list->de_entries[i];
4116
4117                 j = __ocfs2_dx_dir_hash_idx(osb,
4118                                             le32_to_cpu(dx_entry->dx_minor_hash));
4119                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4120
4121                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4122
4123                 /* Each leaf has been passed to the journal already
4124                  * via __ocfs2_dx_dir_new_cluster() */
4125         }
4126
4127         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4128         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4129                offsetof(struct ocfs2_dx_root_block, dr_list));
4130         dx_root->dr_list.l_count =
4131                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4132
4133         /* This should never fail considering we start with an empty
4134          * dx_root. */
4135         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4136         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4137         if (ret)
4138                 mlog_errno(ret);
4139         did_quota = 0;
4140
4141         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4142         ocfs2_journal_dirty(handle, dx_root_bh);
4143
4144 out_commit:
4145         if (ret < 0 && did_quota)
4146                 dquot_free_space_nodirty(dir,
4147                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4148
4149         ocfs2_commit_trans(osb, handle);
4150
4151 out:
4152         if (data_ac)
4153                 ocfs2_free_alloc_context(data_ac);
4154
4155         if (dx_leaves) {
4156                 for (i = 0; i < num_dx_leaves; i++)
4157                         brelse(dx_leaves[i]);
4158                 kfree(dx_leaves);
4159         }
4160         return ret;
4161 }
4162
4163 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4164 {
4165         struct ocfs2_dx_root_block *dx_root;
4166         struct ocfs2_dx_entry_list *entry_list;
4167
4168         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4169         entry_list = &dx_root->dr_entries;
4170
4171         if (le16_to_cpu(entry_list->de_num_used) >=
4172             le16_to_cpu(entry_list->de_count))
4173                 return -ENOSPC;
4174
4175         return 0;
4176 }
4177
4178 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4179                                            struct buffer_head *di_bh,
4180                                            const char *name,
4181                                            int namelen,
4182                                            struct ocfs2_dir_lookup_result *lookup)
4183 {
4184         int ret, free_dx_root = 1;
4185         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4186         struct buffer_head *dx_root_bh = NULL;
4187         struct buffer_head *leaf_bh = NULL;
4188         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4189         struct ocfs2_dx_root_block *dx_root;
4190
4191         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4192         if (ret) {
4193                 mlog_errno(ret);
4194                 goto out;
4195         }
4196
4197         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4198         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4199                 ret = -ENOSPC;
4200                 mlog_errno(ret);
4201                 goto out;
4202         }
4203
4204         if (ocfs2_dx_root_inline(dx_root)) {
4205                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4206
4207                 if (ret == 0)
4208                         goto search_el;
4209
4210                 /*
4211                  * We ran out of room in the root block. Expand it to
4212                  * an extent, then allow ocfs2_find_dir_space_dx to do
4213                  * the rest.
4214                  */
4215                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4216                 if (ret) {
4217                         mlog_errno(ret);
4218                         goto out;
4219                 }
4220         }
4221
4222         /*
4223          * Insert preparation for an indexed directory is split into two
4224          * steps. The call to find_dir_space_dx reserves room in the index for
4225          * an additional item. If we run out of space there, it's a real error
4226          * we can't continue on.
4227          */
4228         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4229                                       namelen, lookup);
4230         if (ret) {
4231                 mlog_errno(ret);
4232                 goto out;
4233         }
4234
4235 search_el:
4236         /*
4237          * Next, we need to find space in the unindexed tree. This call
4238          * searches using the free space linked list. If the unindexed tree
4239          * lacks sufficient space, we'll expand it below. The expansion code
4240          * is smart enough to add any new blocks to the free space list.
4241          */
4242         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4243         if (ret && ret != -ENOSPC) {
4244                 mlog_errno(ret);
4245                 goto out;
4246         }
4247
4248         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4249         lookup->dl_dx_root_bh = dx_root_bh;
4250         free_dx_root = 0;
4251
4252         if (ret == -ENOSPC) {
4253                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4254
4255                 if (ret) {
4256                         mlog_errno(ret);
4257                         goto out;
4258                 }
4259
4260                 /*
4261                  * We make the assumption here that new leaf blocks are added
4262                  * to the front of our free list.
4263                  */
4264                 lookup->dl_prev_leaf_bh = NULL;
4265                 lookup->dl_leaf_bh = leaf_bh;
4266         }
4267
4268 out:
4269         if (free_dx_root)
4270                 brelse(dx_root_bh);
4271         return ret;
4272 }
4273
4274 /*
4275  * Get a directory ready for insert. Any directory allocation required
4276  * happens here. Success returns zero, and enough context in the dir
4277  * lookup result that ocfs2_add_entry() will be able complete the task
4278  * with minimal performance impact.
4279  */
4280 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4281                                  struct inode *dir,
4282                                  struct buffer_head *parent_fe_bh,
4283                                  const char *name,
4284                                  int namelen,
4285                                  struct ocfs2_dir_lookup_result *lookup)
4286 {
4287         int ret;
4288         unsigned int blocks_wanted = 1;
4289         struct buffer_head *bh = NULL;
4290
4291         trace_ocfs2_prepare_dir_for_insert(
4292                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4293
4294         if (!namelen) {
4295                 ret = -EINVAL;
4296                 mlog_errno(ret);
4297                 goto out;
4298         }
4299
4300         /*
4301          * Do this up front to reduce confusion.
4302          *
4303          * The directory might start inline, then be turned into an
4304          * indexed one, in which case we'd need to hash deep inside
4305          * ocfs2_find_dir_space_id(). Since
4306          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4307          * done, there seems no point in spreading out the calls. We
4308          * can optimize away the case where the file system doesn't
4309          * support indexing.
4310          */
4311         if (ocfs2_supports_indexed_dirs(osb))
4312                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4313
4314         if (ocfs2_dir_indexed(dir)) {
4315                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4316                                                       name, namelen, lookup);
4317                 if (ret)
4318                         mlog_errno(ret);
4319                 goto out;
4320         }
4321
4322         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4323                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4324                                               namelen, &bh, &blocks_wanted);
4325         } else
4326                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4327
4328         if (ret && ret != -ENOSPC) {
4329                 mlog_errno(ret);
4330                 goto out;
4331         }
4332
4333         if (ret == -ENOSPC) {
4334                 /*
4335                  * We have to expand the directory to add this name.
4336                  */
4337                 BUG_ON(bh);
4338
4339                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4340                                        lookup, &bh);
4341                 if (ret) {
4342                         if (ret != -ENOSPC)
4343                                 mlog_errno(ret);
4344                         goto out;
4345                 }
4346
4347                 BUG_ON(!bh);
4348         }
4349
4350         lookup->dl_leaf_bh = bh;
4351         bh = NULL;
4352 out:
4353         brelse(bh);
4354         return ret;
4355 }
4356
4357 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4358                                      struct buffer_head *di_bh,
4359                                      struct buffer_head *dx_root_bh)
4360 {
4361         int ret;
4362         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4363         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4364         struct ocfs2_dx_root_block *dx_root;
4365         struct inode *dx_alloc_inode = NULL;
4366         struct buffer_head *dx_alloc_bh = NULL;
4367         handle_t *handle;
4368         u64 blk;
4369         u16 bit;
4370         u64 bg_blkno;
4371
4372         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4373
4374         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4375                                         EXTENT_ALLOC_SYSTEM_INODE,
4376                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4377         if (!dx_alloc_inode) {
4378                 ret = -ENOMEM;
4379                 mlog_errno(ret);
4380                 goto out;
4381         }
4382         mutex_lock(&dx_alloc_inode->i_mutex);
4383
4384         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4385         if (ret) {
4386                 mlog_errno(ret);
4387                 goto out_mutex;
4388         }
4389
4390         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4391         if (IS_ERR(handle)) {
4392                 ret = PTR_ERR(handle);
4393                 mlog_errno(ret);
4394                 goto out_unlock;
4395         }
4396
4397         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4398                                       OCFS2_JOURNAL_ACCESS_WRITE);
4399         if (ret) {
4400                 mlog_errno(ret);
4401                 goto out_commit;
4402         }
4403
4404         spin_lock(&OCFS2_I(dir)->ip_lock);
4405         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4406         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4407         spin_unlock(&OCFS2_I(dir)->ip_lock);
4408         di->i_dx_root = cpu_to_le64(0ULL);
4409         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4410
4411         ocfs2_journal_dirty(handle, di_bh);
4412
4413         blk = le64_to_cpu(dx_root->dr_blkno);
4414         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4415         if (dx_root->dr_suballoc_loc)
4416                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4417         else
4418                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4419         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4420                                        bit, bg_blkno, 1);
4421         if (ret)
4422                 mlog_errno(ret);
4423
4424 out_commit:
4425         ocfs2_commit_trans(osb, handle);
4426
4427 out_unlock:
4428         ocfs2_inode_unlock(dx_alloc_inode, 1);
4429
4430 out_mutex:
4431         mutex_unlock(&dx_alloc_inode->i_mutex);
4432         brelse(dx_alloc_bh);
4433 out:
4434         iput(dx_alloc_inode);
4435         return ret;
4436 }
4437
4438 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4439 {
4440         int ret;
4441         unsigned int uninitialized_var(clen);
4442         u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4443         u64 uninitialized_var(blkno);
4444         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4445         struct buffer_head *dx_root_bh = NULL;
4446         struct ocfs2_dx_root_block *dx_root;
4447         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4448         struct ocfs2_cached_dealloc_ctxt dealloc;
4449         struct ocfs2_extent_tree et;
4450
4451         ocfs2_init_dealloc_ctxt(&dealloc);
4452
4453         if (!ocfs2_dir_indexed(dir))
4454                 return 0;
4455
4456         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4457         if (ret) {
4458                 mlog_errno(ret);
4459                 goto out;
4460         }
4461         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4462
4463         if (ocfs2_dx_root_inline(dx_root))
4464                 goto remove_index;
4465
4466         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4467
4468         /* XXX: What if dr_clusters is too large? */
4469         while (le32_to_cpu(dx_root->dr_clusters)) {
4470                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4471                                               major_hash, &cpos, &blkno, &clen);
4472                 if (ret) {
4473                         mlog_errno(ret);
4474                         goto out;
4475                 }
4476
4477                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4478
4479                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4480                                                &dealloc, 0);
4481                 if (ret) {
4482                         mlog_errno(ret);
4483                         goto out;
4484                 }
4485
4486                 if (cpos == 0)
4487                         break;
4488
4489                 major_hash = cpos - 1;
4490         }
4491
4492 remove_index:
4493         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4494         if (ret) {
4495                 mlog_errno(ret);
4496                 goto out;
4497         }
4498
4499         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4500 out:
4501         ocfs2_schedule_truncate_log_flush(osb, 1);
4502         ocfs2_run_deallocs(osb, &dealloc);
4503
4504         brelse(dx_root_bh);
4505         return ret;
4506 }