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