fs/ocfs2/dir.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * dir.c
   4  *
   5  * Creates, reads, walks and deletes directory-nodes
   6  *
   7  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   8  *
   9  *  Portions of this code from linux/fs/ext3/dir.c
  10  *
  11  *  Copyright (C) 1992, 1993, 1994, 1995
  12  *  Remy Card (card@masi.ibp.fr)
  13  *  Laboratoire MASI - Institut Blaise pascal
  14  *  Universite Pierre et Marie Curie (Paris VI)
  15  *
  16  *   from
  17  *
  18  *   linux/fs/minix/dir.c
  19  *
  20  *   Copyright (C) 1991, 1992 Linus Torvalds
  21  */
  22
  23 #include <linux/fs.h>
  24 #include <linux/types.h>
  25 #include <linux/slab.h>
  26 #include <linux/highmem.h>
  27 #include <linux/quotaops.h>
  28 #include <linux/sort.h>
  29 #include <linux/iversion.h>
  30
  31 #include <cluster/masklog.h>
  32
  33 #include "ocfs2.h"
  34
  35 #include "alloc.h"
  36 #include "blockcheck.h"
  37 #include "dir.h"
  38 #include "dlmglue.h"
  39 #include "extent_map.h"
  40 #include "file.h"
  41 #include "inode.h"
  42 #include "journal.h"
  43 #include "namei.h"
  44 #include "suballoc.h"
  45 #include "super.h"
  46 #include "sysfile.h"
  47 #include "uptodate.h"
  48 #include "ocfs2_trace.h"
  49
  50 #include "buffer_head_io.h"
  51
  52 #define NAMEI_RA_CHUNKS  2
  53 #define NAMEI_RA_BLOCKS  4
  54 #define NAMEI_RA_SIZE        (NAMEI_RA_CHUNKS * NAMEI_RA_BLOCKS)
  55
  56 static int ocfs2_do_extend_dir(struct super_block *sb,
  57                                handle_t *handle,
  58                                struct inode *dir,
  59                                struct buffer_head *parent_fe_bh,
  60                                struct ocfs2_alloc_context *data_ac,
  61                                struct ocfs2_alloc_context *meta_ac,
  62                                struct buffer_head **new_bh);
  63 static int ocfs2_dir_indexed(struct inode *inode);
  64
  65 /*
  66  * These are distinct checks because future versions of the file system will
  67  * want to have a trailing dirent structure independent of indexing.
  68  */
  69 static int ocfs2_supports_dir_trailer(struct inode *dir)
  70 {
  71         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
  72
  73         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
  74                 return 0;
  75
  76         return ocfs2_meta_ecc(osb) || ocfs2_dir_indexed(dir);
  77 }
  78
  79 /*
  80  * "new' here refers to the point at which we're creating a new
  81  * directory via "mkdir()", but also when we're expanding an inline
  82  * directory. In either case, we don't yet have the indexing bit set
  83  * on the directory, so the standard checks will fail in when metaecc
  84  * is turned off. Only directory-initialization type functions should
  85  * use this then. Everything else wants ocfs2_supports_dir_trailer()
  86  */
  87 static int ocfs2_new_dir_wants_trailer(struct inode *dir)
  88 {
  89         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
  90
  91         return ocfs2_meta_ecc(osb) ||
  92                 ocfs2_supports_indexed_dirs(osb);
  93 }
  94
  95 static inline unsigned int ocfs2_dir_trailer_blk_off(struct super_block *sb)
  96 {
  97         return sb->s_blocksize - sizeof(struct ocfs2_dir_block_trailer);
  98 }
  99
 100 #define ocfs2_trailer_from_bh(_bh, _sb) ((struct ocfs2_dir_block_trailer *) ((_bh)->b_data + ocfs2_dir_trailer_blk_off((_sb))))
 101
 102 /* XXX ocfs2_block_dqtrailer() is similar but not quite - can we make
 103  * them more consistent? */
 104 struct ocfs2_dir_block_trailer *ocfs2_dir_trailer_from_size(int blocksize,
 105                                                             void *data)
 106 {
 107         char *p = data;
 108
 109         p += blocksize - sizeof(struct ocfs2_dir_block_trailer);
 110         return (struct ocfs2_dir_block_trailer *)p;
 111 }
 112
 113 /*
 114  * XXX: This is executed once on every dirent. We should consider optimizing
 115  * it.
 116  */
 117 static int ocfs2_skip_dir_trailer(struct inode *dir,
 118                                   struct ocfs2_dir_entry *de,
 119                                   unsigned long offset,
 120                                   unsigned long blklen)
 121 {
 122         unsigned long toff = blklen - sizeof(struct ocfs2_dir_block_trailer);
 123
 124         if (!ocfs2_supports_dir_trailer(dir))
 125                 return 0;
 126
 127         if (offset != toff)
 128                 return 0;
 129
 130         return 1;
 131 }
 132
 133 static void ocfs2_init_dir_trailer(struct inode *inode,
 134                                    struct buffer_head *bh, u16 rec_len)
 135 {
 136         struct ocfs2_dir_block_trailer *trailer;
 137
 138         trailer = ocfs2_trailer_from_bh(bh, inode->i_sb);
 139         strcpy(trailer->db_signature, OCFS2_DIR_TRAILER_SIGNATURE);
 140         trailer->db_compat_rec_len =
 141                         cpu_to_le16(sizeof(struct ocfs2_dir_block_trailer));
 142         trailer->db_parent_dinode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
 143         trailer->db_blkno = cpu_to_le64(bh->b_blocknr);
 144         trailer->db_free_rec_len = cpu_to_le16(rec_len);
 145 }
 146 /*
 147  * Link an unindexed block with a dir trailer structure into the index free
 148  * list. This function will modify dirdata_bh, but assumes you've already
 149  * passed it to the journal.
 150  */
 151 static int ocfs2_dx_dir_link_trailer(struct inode *dir, handle_t *handle,
 152                                      struct buffer_head *dx_root_bh,
 153                                      struct buffer_head *dirdata_bh)
 154 {
 155         int ret;
 156         struct ocfs2_dx_root_block *dx_root;
 157         struct ocfs2_dir_block_trailer *trailer;
 158
 159         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
 160                                       OCFS2_JOURNAL_ACCESS_WRITE);
 161         if (ret) {
 162                 mlog_errno(ret);
 163                 goto out;
 164         }
 165         trailer = ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
 166         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
 167
 168         trailer->db_free_next = dx_root->dr_free_blk;
 169         dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
 170
 171         ocfs2_journal_dirty(handle, dx_root_bh);
 172
 173 out:
 174         return ret;
 175 }
 176
 177 static int ocfs2_free_list_at_root(struct ocfs2_dir_lookup_result *res)
 178 {
 179         return res->dl_prev_leaf_bh == NULL;
 180 }
 181
 182 void ocfs2_free_dir_lookup_result(struct ocfs2_dir_lookup_result *res)
 183 {
 184         brelse(res->dl_dx_root_bh);
 185         brelse(res->dl_leaf_bh);
 186         brelse(res->dl_dx_leaf_bh);
 187         brelse(res->dl_prev_leaf_bh);
 188 }
 189
 190 static int ocfs2_dir_indexed(struct inode *inode)
 191 {
 192         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INDEXED_DIR_FL)
 193                 return 1;
 194         return 0;
 195 }
 196
 197 static inline int ocfs2_dx_root_inline(struct ocfs2_dx_root_block *dx_root)
 198 {
 199         return dx_root->dr_flags & OCFS2_DX_FLAG_INLINE;
 200 }
 201
 202 /*
 203  * Hashing code adapted from ext3
 204  */
 205 #define DELTA 0x9E3779B9
 206
 207 static void TEA_transform(__u32 buf[4], __u32 const in[])
 208 {
 209         __u32   sum = 0;
 210         __u32   b0 = buf[0], b1 = buf[1];
 211         __u32   a = in[0], b = in[1], c = in[2], d = in[3];
 212         int     n = 16;
 213
 214         do {
 215                 sum += DELTA;
 216                 b0 += ((b1 << 4)+a) ^ (b1+sum) ^ ((b1 >> 5)+b);
 217                 b1 += ((b0 << 4)+c) ^ (b0+sum) ^ ((b0 >> 5)+d);
 218         } while (--n);
 219
 220         buf[0] += b0;
 221         buf[1] += b1;
 222 }
 223
 224 static void str2hashbuf(const char *msg, int len, __u32 *buf, int num)
 225 {
 226         __u32   pad, val;
 227         int     i;
 228
 229         pad = (__u32)len | ((__u32)len << 8);
 230         pad |= pad << 16;
 231
 232         val = pad;
 233         if (len > num*4)
 234                 len = num * 4;
 235         for (i = 0; i < len; i++) {
 236                 if ((i % 4) == 0)
 237                         val = pad;
 238                 val = msg[i] + (val << 8);
 239                 if ((i % 4) == 3) {
 240                         *buf++ = val;
 241                         val = pad;
 242                         num--;
 243                 }
 244         }
 245         if (--num >= 0)
 246                 *buf++ = val;
 247         while (--num >= 0)
 248                 *buf++ = pad;
 249 }
 250
 251 static void ocfs2_dx_dir_name_hash(struct inode *dir, const char *name, int len,
 252                                    struct ocfs2_dx_hinfo *hinfo)
 253 {
 254         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
 255         const char      *p;
 256         __u32           in[8], buf[4];
 257
 258         /*
 259          * XXX: Is this really necessary, if the index is never looked
 260          * at by readdir? Is a hash value of '0' a bad idea?
 261          */
 262         if ((len == 1 && !strncmp(".", name, 1)) ||
 263             (len == 2 && !strncmp("..", name, 2))) {
 264                 buf[0] = buf[1] = 0;
 265                 goto out;
 266         }
 267
 268 #ifdef OCFS2_DEBUG_DX_DIRS
 269         /*
 270          * This makes it very easy to debug indexing problems. We
 271          * should never allow this to be selected without hand editing
 272          * this file though.
 273          */
 274         buf[0] = buf[1] = len;
 275         goto out;
 276 #endif
 277
 278         memcpy(buf, osb->osb_dx_seed, sizeof(buf));
 279
 280         p = name;
 281         while (len > 0) {
 282                 str2hashbuf(p, len, in, 4);
 283                 TEA_transform(buf, in);
 284                 len -= 16;
 285                 p += 16;
 286         }
 287
 288 out:
 289         hinfo->major_hash = buf[0];
 290         hinfo->minor_hash = buf[1];
 291 }
 292
 293 /*
 294  * bh passed here can be an inode block or a dir data block, depending
 295  * on the inode inline data flag.
 296  */
 297 static int ocfs2_check_dir_entry(struct inode * dir,
 298                                  struct ocfs2_dir_entry * de,
 299                                  struct buffer_head * bh,
 300                                  unsigned long offset)
 301 {
 302         const char *error_msg = NULL;
 303         const int rlen = le16_to_cpu(de->rec_len);
 304
 305         if (unlikely(rlen < OCFS2_DIR_REC_LEN(1)))
 306                 error_msg = "rec_len is smaller than minimal";
 307         else if (unlikely(rlen % 4 != 0))
 308                 error_msg = "rec_len % 4 != 0";
 309         else if (unlikely(rlen < OCFS2_DIR_REC_LEN(de->name_len)))
 310                 error_msg = "rec_len is too small for name_len";
 311         else if (unlikely(
 312                  ((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize))
 313                 error_msg = "directory entry across blocks";
 314
 315         if (unlikely(error_msg != NULL))
 316                 mlog(ML_ERROR, "bad entry in directory #%llu: %s - "
 317                      "offset=%lu, inode=%llu, rec_len=%d, name_len=%d\n",
 318                      (unsigned long long)OCFS2_I(dir)->ip_blkno, error_msg,
 319                      offset, (unsigned long long)le64_to_cpu(de->inode), rlen,
 320                      de->name_len);
 321
 322         return error_msg == NULL ? 1 : 0;
 323 }
 324
 325 static inline int ocfs2_match(int len,
 326                               const char * const name,
 327                               struct ocfs2_dir_entry *de)
 328 {
 329         if (len != de->name_len)
 330                 return 0;
 331         if (!de->inode)
 332                 return 0;
 333         return !memcmp(name, de->name, len);
 334 }
 335
 336 /*
 337  * Returns 0 if not found, -1 on failure, and 1 on success
 338  */
 339 static inline int ocfs2_search_dirblock(struct buffer_head *bh,
 340                                         struct inode *dir,
 341                                         const char *name, int namelen,
 342                                         unsigned long offset,
 343                                         char *first_de,
 344                                         unsigned int bytes,
 345                                         struct ocfs2_dir_entry **res_dir)
 346 {
 347         struct ocfs2_dir_entry *de;
 348         char *dlimit, *de_buf;
 349         int de_len;
 350         int ret = 0;
 351
 352         de_buf = first_de;
 353         dlimit = de_buf + bytes;
 354
 355         while (de_buf < dlimit) {
 356                 /* this code is executed quadratically often */
 357                 /* do minimal checking `by hand' */
 358
 359                 de = (struct ocfs2_dir_entry *) de_buf;
 360
 361                 if (de_buf + namelen <= dlimit &&
 362                     ocfs2_match(namelen, name, de)) {
 363                         /* found a match - just to be sure, do a full check */
 364                         if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
 365                                 ret = -1;
 366                                 goto bail;
 367                         }
 368                         *res_dir = de;
 369                         ret = 1;
 370                         goto bail;
 371                 }
 372
 373                 /* prevent looping on a bad block */
 374                 de_len = le16_to_cpu(de->rec_len);
 375                 if (de_len <= 0) {
 376                         ret = -1;
 377                         goto bail;
 378                 }
 379
 380                 de_buf += de_len;
 381                 offset += de_len;
 382         }
 383
 384 bail:
 385         trace_ocfs2_search_dirblock(ret);
 386         return ret;
 387 }
 388
 389 static struct buffer_head *ocfs2_find_entry_id(const char *name,
 390                                                int namelen,
 391                                                struct inode *dir,
 392                                                struct ocfs2_dir_entry **res_dir)
 393 {
 394         int ret, found;
 395         struct buffer_head *di_bh = NULL;
 396         struct ocfs2_dinode *di;
 397         struct ocfs2_inline_data *data;
 398
 399         ret = ocfs2_read_inode_block(dir, &di_bh);
 400         if (ret) {
 401                 mlog_errno(ret);
 402                 goto out;
 403         }
 404
 405         di = (struct ocfs2_dinode *)di_bh->b_data;
 406         data = &di->id2.i_data;
 407
 408         found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
 409                                       data->id_data, i_size_read(dir), res_dir);
 410         if (found == 1)
 411                 return di_bh;
 412
 413         brelse(di_bh);
 414 out:
 415         return NULL;
 416 }
 417
 418 static int ocfs2_validate_dir_block(struct super_block *sb,
 419                                     struct buffer_head *bh)
 420 {
 421         int rc;
 422         struct ocfs2_dir_block_trailer *trailer =
 423                 ocfs2_trailer_from_bh(bh, sb);
 424
 425
 426         /*
 427          * We don't validate dirents here, that's handled
 428          * in-place when the code walks them.
 429          */
 430         trace_ocfs2_validate_dir_block((unsigned long long)bh->b_blocknr);
 431
 432         BUG_ON(!buffer_uptodate(bh));
 433
 434         /*
 435          * If the ecc fails, we return the error but otherwise
 436          * leave the filesystem running.  We know any error is
 437          * local to this block.
 438          *
 439          * Note that we are safe to call this even if the directory
 440          * doesn't have a trailer.  Filesystems without metaecc will do
 441          * nothing, and filesystems with it will have one.
 442          */
 443         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
 444         if (rc)
 445                 mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
 446                      (unsigned long long)bh->b_blocknr);
 447
 448         return rc;
 449 }
 450
 451 /*
 452  * Validate a directory trailer.
 453  *
 454  * We check the trailer here rather than in ocfs2_validate_dir_block()
 455  * because that function doesn't have the inode to test.
 456  */
 457 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
 458 {
 459         int rc = 0;
 460         struct ocfs2_dir_block_trailer *trailer;
 461
 462         trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
 463         if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
 464                 rc = ocfs2_error(dir->i_sb,
 465                                  "Invalid dirblock #%llu: signature = %.*s\n",
 466                                  (unsigned long long)bh->b_blocknr, 7,
 467                                  trailer->db_signature);
 468                 goto out;
 469         }
 470         if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
 471                 rc = ocfs2_error(dir->i_sb,
 472                                  "Directory block #%llu has an invalid db_blkno of %llu\n",
 473                                  (unsigned long long)bh->b_blocknr,
 474                                  (unsigned long long)le64_to_cpu(trailer->db_blkno));
 475                 goto out;
 476         }
 477         if (le64_to_cpu(trailer->db_parent_dinode) !=
 478             OCFS2_I(dir)->ip_blkno) {
 479                 rc = ocfs2_error(dir->i_sb,
 480                                  "Directory block #%llu on dinode #%llu has an invalid parent_dinode of %llu\n",
 481                                  (unsigned long long)bh->b_blocknr,
 482                                  (unsigned long long)OCFS2_I(dir)->ip_blkno,
 483                                  (unsigned long long)le64_to_cpu(trailer->db_blkno));
 484                 goto out;
 485         }
 486 out:
 487         return rc;
 488 }
 489
 490 /*
 491  * This function forces all errors to -EIO for consistency with its
 492  * predecessor, ocfs2_bread().  We haven't audited what returning the
 493  * real error codes would do to callers.  We log the real codes with
 494  * mlog_errno() before we squash them.
 495  */
 496 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
 497                                 struct buffer_head **bh, int flags)
 498 {
 499         int rc = 0;
 500         struct buffer_head *tmp = *bh;
 501
 502         rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
 503                                     ocfs2_validate_dir_block);
 504         if (rc) {
 505                 mlog_errno(rc);
 506                 goto out;
 507         }
 508
 509         if (!(flags & OCFS2_BH_READAHEAD) &&
 510             ocfs2_supports_dir_trailer(inode)) {
 511                 rc = ocfs2_check_dir_trailer(inode, tmp);
 512                 if (rc) {
 513                         if (!*bh)
 514                                 brelse(tmp);
 515                         mlog_errno(rc);
 516                         goto out;
 517                 }
 518         }
 519
 520         /* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
 521         if (!*bh)
 522                 *bh = tmp;
 523
 524 out:
 525         return rc ? -EIO : 0;
 526 }
 527
 528 /*
 529  * Read the block at 'phys' which belongs to this directory
 530  * inode. This function does no virtual->physical block translation -
 531  * what's passed in is assumed to be a valid directory block.
 532  */
 533 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
 534                                        struct buffer_head **bh)
 535 {
 536         int ret;
 537         struct buffer_head *tmp = *bh;
 538
 539         ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
 540                                ocfs2_validate_dir_block);
 541         if (ret) {
 542                 mlog_errno(ret);
 543                 goto out;
 544         }
 545
 546         if (ocfs2_supports_dir_trailer(dir)) {
 547                 ret = ocfs2_check_dir_trailer(dir, tmp);
 548                 if (ret) {
 549                         if (!*bh)
 550                                 brelse(tmp);
 551                         mlog_errno(ret);
 552                         goto out;
 553                 }
 554         }
 555
 556         if (!ret && !*bh)
 557                 *bh = tmp;
 558 out:
 559         return ret;
 560 }
 561
 562 static int ocfs2_validate_dx_root(struct super_block *sb,
 563                                   struct buffer_head *bh)
 564 {
 565         int ret;
 566         struct ocfs2_dx_root_block *dx_root;
 567
 568         BUG_ON(!buffer_uptodate(bh));
 569
 570         dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
 571
 572         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
 573         if (ret) {
 574                 mlog(ML_ERROR,
 575                      "Checksum failed for dir index root block %llu\n",
 576                      (unsigned long long)bh->b_blocknr);
 577                 return ret;
 578         }
 579
 580         if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
 581                 ret = ocfs2_error(sb,
 582                                   "Dir Index Root # %llu has bad signature %.*s\n",
 583                                   (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
 584                                   7, dx_root->dr_signature);
 585         }
 586
 587         return ret;
 588 }
 589
 590 static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
 591                               struct buffer_head **dx_root_bh)
 592 {
 593         int ret;
 594         u64 blkno = le64_to_cpu(di->i_dx_root);
 595         struct buffer_head *tmp = *dx_root_bh;
 596
 597         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 598                                ocfs2_validate_dx_root);
 599
 600         /* If ocfs2_read_block() got us a new bh, pass it up. */
 601         if (!ret && !*dx_root_bh)
 602                 *dx_root_bh = tmp;
 603
 604         return ret;
 605 }
 606
 607 static int ocfs2_validate_dx_leaf(struct super_block *sb,
 608                                   struct buffer_head *bh)
 609 {
 610         int ret;
 611         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
 612
 613         BUG_ON(!buffer_uptodate(bh));
 614
 615         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
 616         if (ret) {
 617                 mlog(ML_ERROR,
 618                      "Checksum failed for dir index leaf block %llu\n",
 619                      (unsigned long long)bh->b_blocknr);
 620                 return ret;
 621         }
 622
 623         if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
 624                 ret = ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s\n",
 625                                   7, dx_leaf->dl_signature);
 626         }
 627
 628         return ret;
 629 }
 630
 631 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
 632                               struct buffer_head **dx_leaf_bh)
 633 {
 634         int ret;
 635         struct buffer_head *tmp = *dx_leaf_bh;
 636
 637         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 638                                ocfs2_validate_dx_leaf);
 639
 640         /* If ocfs2_read_block() got us a new bh, pass it up. */
 641         if (!ret && !*dx_leaf_bh)
 642                 *dx_leaf_bh = tmp;
 643
 644         return ret;
 645 }
 646
 647 /*
 648  * Read a series of dx_leaf blocks. This expects all buffer_head
 649  * pointers to be NULL on function entry.
 650  */
 651 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
 652                                 struct buffer_head **dx_leaf_bhs)
 653 {
 654         int ret;
 655
 656         ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
 657                                 ocfs2_validate_dx_leaf);
 658         if (ret)
 659                 mlog_errno(ret);
 660
 661         return ret;
 662 }
 663
 664 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
 665                                                struct inode *dir,
 666                                                struct ocfs2_dir_entry **res_dir)
 667 {
 668         struct super_block *sb;
 669         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 670         struct buffer_head *bh, *ret = NULL;
 671         unsigned long start, block, b;
 672         int ra_max = 0;         /* Number of bh's in the readahead
 673                                    buffer, bh_use[] */
 674         int ra_ptr = 0;         /* Current index into readahead
 675                                    buffer */
 676         int num = 0;
 677         int nblocks, i;
 678
 679         sb = dir->i_sb;
 680
 681         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 682         start = OCFS2_I(dir)->ip_dir_start_lookup;
 683         if (start >= nblocks)
 684                 start = 0;
 685         block = start;
 686
 687 restart:
 688         do {
 689                 /*
 690                  * We deal with the read-ahead logic here.
 691                  */
 692                 if (ra_ptr >= ra_max) {
 693                         /* Refill the readahead buffer */
 694                         ra_ptr = 0;
 695                         b = block;
 696                         for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
 697                                 /*
 698                                  * Terminate if we reach the end of the
 699                                  * directory and must wrap, or if our
 700                                  * search has finished at this block.
 701                                  */
 702                                 if (b >= nblocks || (num && block == start)) {
 703                                         bh_use[ra_max] = NULL;
 704                                         break;
 705                                 }
 706                                 num++;
 707
 708                                 bh = NULL;
 709                                 ocfs2_read_dir_block(dir, b++, &bh,
 710                                                            OCFS2_BH_READAHEAD);
 711                                 bh_use[ra_max] = bh;
 712                         }
 713                 }
 714                 if ((bh = bh_use[ra_ptr++]) == NULL)
 715                         goto next;
 716                 if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
 717                         /* read error, skip block & hope for the best.
 718                          * ocfs2_read_dir_block() has released the bh. */
 719                         mlog(ML_ERROR, "reading directory %llu, "
 720                                     "offset %lu\n",
 721                                     (unsigned long long)OCFS2_I(dir)->ip_blkno,
 722                                     block);
 723                         goto next;
 724                 }
 725                 i = ocfs2_search_dirblock(bh, dir, name, namelen,
 726                                           block << sb->s_blocksize_bits,
 727                                           bh->b_data, sb->s_blocksize,
 728                                           res_dir);
 729                 if (i == 1) {
 730                         OCFS2_I(dir)->ip_dir_start_lookup = block;
 731                         ret = bh;
 732                         goto cleanup_and_exit;
 733                 } else {
 734                         brelse(bh);
 735                         if (i < 0)
 736                                 goto cleanup_and_exit;
 737                 }
 738         next:
 739                 if (++block >= nblocks)
 740                         block = 0;
 741         } while (block != start);
 742
 743         /*
 744          * If the directory has grown while we were searching, then
 745          * search the last part of the directory before giving up.
 746          */
 747         block = nblocks;
 748         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 749         if (block < nblocks) {
 750                 start = 0;
 751                 goto restart;
 752         }
 753
 754 cleanup_and_exit:
 755         /* Clean up the read-ahead blocks */
 756         for (; ra_ptr < ra_max; ra_ptr++)
 757                 brelse(bh_use[ra_ptr]);
 758
 759         trace_ocfs2_find_entry_el(ret);
 760         return ret;
 761 }
 762
 763 static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
 764                                    struct ocfs2_extent_list *el,
 765                                    u32 major_hash,
 766                                    u32 *ret_cpos,
 767                                    u64 *ret_phys_blkno,
 768                                    unsigned int *ret_clen)
 769 {
 770         int ret = 0, i, found;
 771         struct buffer_head *eb_bh = NULL;
 772         struct ocfs2_extent_block *eb;
 773         struct ocfs2_extent_rec *rec = NULL;
 774
 775         if (el->l_tree_depth) {
 776                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
 777                                       &eb_bh);
 778                 if (ret) {
 779                         mlog_errno(ret);
 780                         goto out;
 781                 }
 782
 783                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 784                 el = &eb->h_list;
 785
 786                 if (el->l_tree_depth) {
 787                         ret = ocfs2_error(inode->i_sb,
 788                                           "Inode %lu has non zero tree depth in btree tree block %llu\n",
 789                                           inode->i_ino,
 790                                           (unsigned long long)eb_bh->b_blocknr);
 791                         goto out;
 792                 }
 793         }
 794
 795         found = 0;
 796         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
 797                 rec = &el->l_recs[i];
 798
 799                 if (le32_to_cpu(rec->e_cpos) <= major_hash) {
 800                         found = 1;
 801                         break;
 802                 }
 803         }
 804
 805         if (!found) {
 806                 ret = ocfs2_error(inode->i_sb,
 807                                   "Inode %lu has bad extent record (%u, %u, 0) in btree\n",
 808                                   inode->i_ino,
 809                                   le32_to_cpu(rec->e_cpos),
 810                                   ocfs2_rec_clusters(el, rec));
 811                 goto out;
 812         }
 813
 814         if (ret_phys_blkno)
 815                 *ret_phys_blkno = le64_to_cpu(rec->e_blkno);
 816         if (ret_cpos)
 817                 *ret_cpos = le32_to_cpu(rec->e_cpos);
 818         if (ret_clen)
 819                 *ret_clen = le16_to_cpu(rec->e_leaf_clusters);
 820
 821 out:
 822         brelse(eb_bh);
 823         return ret;
 824 }
 825
 826 /*
 827  * Returns the block index, from the start of the cluster which this
 828  * hash belongs too.
 829  */
 830 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 831                                                    u32 minor_hash)
 832 {
 833         return minor_hash & osb->osb_dx_mask;
 834 }
 835
 836 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 837                                           struct ocfs2_dx_hinfo *hinfo)
 838 {
 839         return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
 840 }
 841
 842 static int ocfs2_dx_dir_lookup(struct inode *inode,
 843                                struct ocfs2_extent_list *el,
 844                                struct ocfs2_dx_hinfo *hinfo,
 845                                u32 *ret_cpos,
 846                                u64 *ret_phys_blkno)
 847 {
 848         int ret = 0;
 849         unsigned int cend, clen;
 850         u32 cpos;
 851         u64 blkno;
 852         u32 name_hash = hinfo->major_hash;
 853
 854         ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
 855                                       &clen);
 856         if (ret) {
 857                 mlog_errno(ret);
 858                 goto out;
 859         }
 860
 861         cend = cpos + clen;
 862         if (name_hash >= cend) {
 863                 /* We want the last cluster */
 864                 blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
 865                 cpos += clen - 1;
 866         } else {
 867                 blkno += ocfs2_clusters_to_blocks(inode->i_sb,
 868                                                   name_hash - cpos);
 869                 cpos = name_hash;
 870         }
 871
 872         /*
 873          * We now have the cluster which should hold our entry. To
 874          * find the exact block from the start of the cluster to
 875          * search, we take the lower bits of the hash.
 876          */
 877         blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
 878
 879         if (ret_phys_blkno)
 880                 *ret_phys_blkno = blkno;
 881         if (ret_cpos)
 882                 *ret_cpos = cpos;
 883
 884 out:
 885
 886         return ret;
 887 }
 888
 889 static int ocfs2_dx_dir_search(const char *name, int namelen,
 890                                struct inode *dir,
 891                                struct ocfs2_dx_root_block *dx_root,
 892                                struct ocfs2_dir_lookup_result *res)
 893 {
 894         int ret, i, found;
 895         u64 phys;
 896         struct buffer_head *dx_leaf_bh = NULL;
 897         struct ocfs2_dx_leaf *dx_leaf;
 898         struct ocfs2_dx_entry *dx_entry = NULL;
 899         struct buffer_head *dir_ent_bh = NULL;
 900         struct ocfs2_dir_entry *dir_ent = NULL;
 901         struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
 902         struct ocfs2_extent_list *dr_el;
 903         struct ocfs2_dx_entry_list *entry_list;
 904
 905         ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
 906
 907         if (ocfs2_dx_root_inline(dx_root)) {
 908                 entry_list = &dx_root->dr_entries;
 909                 goto search;
 910         }
 911
 912         dr_el = &dx_root->dr_list;
 913
 914         ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
 915         if (ret) {
 916                 mlog_errno(ret);
 917                 goto out;
 918         }
 919
 920         trace_ocfs2_dx_dir_search((unsigned long long)OCFS2_I(dir)->ip_blkno,
 921                                   namelen, name, hinfo->major_hash,
 922                                   hinfo->minor_hash, (unsigned long long)phys);
 923
 924         ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
 925         if (ret) {
 926                 mlog_errno(ret);
 927                 goto out;
 928         }
 929
 930         dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
 931
 932         trace_ocfs2_dx_dir_search_leaf_info(
 933                         le16_to_cpu(dx_leaf->dl_list.de_num_used),
 934                         le16_to_cpu(dx_leaf->dl_list.de_count));
 935
 936         entry_list = &dx_leaf->dl_list;
 937
 938 search:
 939         /*
 940          * Empty leaf is legal, so no need to check for that.
 941          */
 942         found = 0;
 943         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
 944                 dx_entry = &entry_list->de_entries[i];
 945
 946                 if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
 947                     || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
 948                         continue;
 949
 950                 /*
 951                  * Search unindexed leaf block now. We're not
 952                  * guaranteed to find anything.
 953                  */
 954                 ret = ocfs2_read_dir_block_direct(dir,
 955                                           le64_to_cpu(dx_entry->dx_dirent_blk),
 956                                           &dir_ent_bh);
 957                 if (ret) {
 958                         mlog_errno(ret);
 959                         goto out;
 960                 }
 961
 962                 /*
 963                  * XXX: We should check the unindexed block here,
 964                  * before using it.
 965                  */
 966
 967                 found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
 968                                               0, dir_ent_bh->b_data,
 969                                               dir->i_sb->s_blocksize, &dir_ent);
 970                 if (found == 1)
 971                         break;
 972
 973                 if (found == -1) {
 974                         /* This means we found a bad directory entry. */
 975                         ret = -EIO;
 976                         mlog_errno(ret);
 977                         goto out;
 978                 }
 979
 980                 brelse(dir_ent_bh);
 981                 dir_ent_bh = NULL;
 982         }
 983
 984         if (found <= 0) {
 985                 ret = -ENOENT;
 986                 goto out;
 987         }
 988
 989         res->dl_leaf_bh = dir_ent_bh;
 990         res->dl_entry = dir_ent;
 991         res->dl_dx_leaf_bh = dx_leaf_bh;
 992         res->dl_dx_entry = dx_entry;
 993
 994         ret = 0;
 995 out:
 996         if (ret) {
 997                 brelse(dx_leaf_bh);
 998                 brelse(dir_ent_bh);
 999         }
1000         return ret;
1001 }
1002
1003 static int ocfs2_find_entry_dx(const char *name, int namelen,
1004                                struct inode *dir,
1005                                struct ocfs2_dir_lookup_result *lookup)
1006 {
1007         int ret;
1008         struct buffer_head *di_bh = NULL;
1009         struct ocfs2_dinode *di;
1010         struct buffer_head *dx_root_bh = NULL;
1011         struct ocfs2_dx_root_block *dx_root;
1012
1013         ret = ocfs2_read_inode_block(dir, &di_bh);
1014         if (ret) {
1015                 mlog_errno(ret);
1016                 goto out;
1017         }
1018
1019         di = (struct ocfs2_dinode *)di_bh->b_data;
1020
1021         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1022         if (ret) {
1023                 mlog_errno(ret);
1024                 goto out;
1025         }
1026         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1027
1028         ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1029         if (ret) {
1030                 if (ret != -ENOENT)
1031                         mlog_errno(ret);
1032                 goto out;
1033         }
1034
1035         lookup->dl_dx_root_bh = dx_root_bh;
1036         dx_root_bh = NULL;
1037 out:
1038         brelse(di_bh);
1039         brelse(dx_root_bh);
1040         return ret;
1041 }
1042
1043 /*
1044  * Try to find an entry of the provided name within 'dir'.
1045  *
1046  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1047  * returned and the struct 'res' will contain information useful to
1048  * other directory manipulation functions.
1049  *
1050  * Caller can NOT assume anything about the contents of the
1051  * buffer_heads - they are passed back only so that it can be passed
1052  * into any one of the manipulation functions (add entry, delete
1053  * entry, etc). As an example, bh in the extent directory case is a
1054  * data block, in the inline-data case it actually points to an inode,
1055  * in the indexed directory case, multiple buffers are involved.
1056  */
1057 int ocfs2_find_entry(const char *name, int namelen,
1058                      struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1059 {
1060         struct buffer_head *bh;
1061         struct ocfs2_dir_entry *res_dir = NULL;
1062
1063         if (ocfs2_dir_indexed(dir))
1064                 return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1065
1066         /*
1067          * The unindexed dir code only uses part of the lookup
1068          * structure, so there's no reason to push it down further
1069          * than this.
1070          */
1071         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1072                 bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1073         else
1074                 bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1075
1076         if (bh == NULL)
1077                 return -ENOENT;
1078
1079         lookup->dl_leaf_bh = bh;
1080         lookup->dl_entry = res_dir;
1081         return 0;
1082 }
1083
1084 /*
1085  * Update inode number and type of a previously found directory entry.
1086  */
1087 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1088                        struct ocfs2_dir_lookup_result *res,
1089                        struct inode *new_entry_inode)
1090 {
1091         int ret;
1092         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1093         struct ocfs2_dir_entry *de = res->dl_entry;
1094         struct buffer_head *de_bh = res->dl_leaf_bh;
1095
1096         /*
1097          * The same code works fine for both inline-data and extent
1098          * based directories, so no need to split this up.  The only
1099          * difference is the journal_access function.
1100          */
1101
1102         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1103                 access = ocfs2_journal_access_di;
1104
1105         ret = access(handle, INODE_CACHE(dir), de_bh,
1106                      OCFS2_JOURNAL_ACCESS_WRITE);
1107         if (ret) {
1108                 mlog_errno(ret);
1109                 goto out;
1110         }
1111
1112         de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1113         ocfs2_set_de_type(de, new_entry_inode->i_mode);
1114
1115         ocfs2_journal_dirty(handle, de_bh);
1116
1117 out:
1118         return ret;
1119 }
1120
1121 /*
1122  * __ocfs2_delete_entry deletes a directory entry by merging it with the
1123  * previous entry
1124  */
1125 static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1126                                 struct ocfs2_dir_entry *de_del,
1127                                 struct buffer_head *bh, char *first_de,
1128                                 unsigned int bytes)
1129 {
1130         struct ocfs2_dir_entry *de, *pde;
1131         int i, status = -ENOENT;
1132         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1133
1134         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1135                 access = ocfs2_journal_access_di;
1136
1137         i = 0;
1138         pde = NULL;
1139         de = (struct ocfs2_dir_entry *) first_de;
1140         while (i < bytes) {
1141                 if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1142                         status = -EIO;
1143                         mlog_errno(status);
1144                         goto bail;
1145                 }
1146                 if (de == de_del)  {
1147                         status = access(handle, INODE_CACHE(dir), bh,
1148                                         OCFS2_JOURNAL_ACCESS_WRITE);
1149                         if (status < 0) {
1150                                 status = -EIO;
1151                                 mlog_errno(status);
1152                                 goto bail;
1153                         }
1154                         if (pde)
1155                                 le16_add_cpu(&pde->rec_len,
1156                                                 le16_to_cpu(de->rec_len));
1157                         de->inode = 0;
1158                         inode_inc_iversion(dir);
1159                         ocfs2_journal_dirty(handle, bh);
1160                         goto bail;
1161                 }
1162                 i += le16_to_cpu(de->rec_len);
1163                 pde = de;
1164                 de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1165         }
1166 bail:
1167         return status;
1168 }
1169
1170 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1171 {
1172         unsigned int hole;
1173
1174         if (le64_to_cpu(de->inode) == 0)
1175                 hole = le16_to_cpu(de->rec_len);
1176         else
1177                 hole = le16_to_cpu(de->rec_len) -
1178                         OCFS2_DIR_REC_LEN(de->name_len);
1179
1180         return hole;
1181 }
1182
1183 static int ocfs2_find_max_rec_len(struct super_block *sb,
1184                                   struct buffer_head *dirblock_bh)
1185 {
1186         int size, this_hole, largest_hole = 0;
1187         char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1188         struct ocfs2_dir_entry *de;
1189
1190         trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1191         size = ocfs2_dir_trailer_blk_off(sb);
1192         limit = start + size;
1193         de_buf = start;
1194         de = (struct ocfs2_dir_entry *)de_buf;
1195         do {
1196                 if (de_buf != trailer) {
1197                         this_hole = ocfs2_figure_dirent_hole(de);
1198                         if (this_hole > largest_hole)
1199                                 largest_hole = this_hole;
1200                 }
1201
1202                 de_buf += le16_to_cpu(de->rec_len);
1203                 de = (struct ocfs2_dir_entry *)de_buf;
1204         } while (de_buf < limit);
1205
1206         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1207                 return largest_hole;
1208         return 0;
1209 }
1210
1211 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1212                                        int index)
1213 {
1214         int num_used = le16_to_cpu(entry_list->de_num_used);
1215
1216         if (num_used == 1 || index == (num_used - 1))
1217                 goto clear;
1218
1219         memmove(&entry_list->de_entries[index],
1220                 &entry_list->de_entries[index + 1],
1221                 (num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1222 clear:
1223         num_used--;
1224         memset(&entry_list->de_entries[num_used], 0,
1225                sizeof(struct ocfs2_dx_entry));
1226         entry_list->de_num_used = cpu_to_le16(num_used);
1227 }
1228
1229 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1230                                  struct ocfs2_dir_lookup_result *lookup)
1231 {
1232         int ret, index, max_rec_len, add_to_free_list = 0;
1233         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1234         struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1235         struct ocfs2_dx_leaf *dx_leaf;
1236         struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1237         struct ocfs2_dir_block_trailer *trailer;
1238         struct ocfs2_dx_root_block *dx_root;
1239         struct ocfs2_dx_entry_list *entry_list;
1240
1241         /*
1242          * This function gets a bit messy because we might have to
1243          * modify the root block, regardless of whether the indexed
1244          * entries are stored inline.
1245          */
1246
1247         /*
1248          * *Only* set 'entry_list' here, based on where we're looking
1249          * for the indexed entries. Later, we might still want to
1250          * journal both blocks, based on free list state.
1251          */
1252         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1253         if (ocfs2_dx_root_inline(dx_root)) {
1254                 entry_list = &dx_root->dr_entries;
1255         } else {
1256                 dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1257                 entry_list = &dx_leaf->dl_list;
1258         }
1259
1260         /* Neither of these are a disk corruption - that should have
1261          * been caught by lookup, before we got here. */
1262         BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1263         BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1264
1265         index = (char *)dx_entry - (char *)entry_list->de_entries;
1266         index /= sizeof(*dx_entry);
1267
1268         if (index >= le16_to_cpu(entry_list->de_num_used)) {
1269                 mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1270                      (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1271                      entry_list, dx_entry);
1272                 return -EIO;
1273         }
1274
1275         /*
1276          * We know that removal of this dirent will leave enough room
1277          * for a new one, so add this block to the free list if it
1278          * isn't already there.
1279          */
1280         trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1281         if (trailer->db_free_rec_len == 0)
1282                 add_to_free_list = 1;
1283
1284         /*
1285          * Add the block holding our index into the journal before
1286          * removing the unindexed entry. If we get an error return
1287          * from __ocfs2_delete_entry(), then it hasn't removed the
1288          * entry yet. Likewise, successful return means we *must*
1289          * remove the indexed entry.
1290          *
1291          * We're also careful to journal the root tree block here as
1292          * the entry count needs to be updated. Also, we might be
1293          * adding to the start of the free list.
1294          */
1295         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1296                                       OCFS2_JOURNAL_ACCESS_WRITE);
1297         if (ret) {
1298                 mlog_errno(ret);
1299                 goto out;
1300         }
1301
1302         if (!ocfs2_dx_root_inline(dx_root)) {
1303                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1304                                               lookup->dl_dx_leaf_bh,
1305                                               OCFS2_JOURNAL_ACCESS_WRITE);
1306                 if (ret) {
1307                         mlog_errno(ret);
1308                         goto out;
1309                 }
1310         }
1311
1312         trace_ocfs2_delete_entry_dx((unsigned long long)OCFS2_I(dir)->ip_blkno,
1313                                     index);
1314
1315         ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1316                                    leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1317         if (ret) {
1318                 mlog_errno(ret);
1319                 goto out;
1320         }
1321
1322         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1323         trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1324         if (add_to_free_list) {
1325                 trailer->db_free_next = dx_root->dr_free_blk;
1326                 dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1327                 ocfs2_journal_dirty(handle, dx_root_bh);
1328         }
1329
1330         /* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1331         ocfs2_journal_dirty(handle, leaf_bh);
1332
1333         le32_add_cpu(&dx_root->dr_num_entries, -1);
1334         ocfs2_journal_dirty(handle, dx_root_bh);
1335
1336         ocfs2_dx_list_remove_entry(entry_list, index);
1337
1338         if (!ocfs2_dx_root_inline(dx_root))
1339                 ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1340
1341 out:
1342         return ret;
1343 }
1344
1345 static inline int ocfs2_delete_entry_id(handle_t *handle,
1346                                         struct inode *dir,
1347                                         struct ocfs2_dir_entry *de_del,
1348                                         struct buffer_head *bh)
1349 {
1350         int ret;
1351         struct buffer_head *di_bh = NULL;
1352         struct ocfs2_dinode *di;
1353         struct ocfs2_inline_data *data;
1354
1355         ret = ocfs2_read_inode_block(dir, &di_bh);
1356         if (ret) {
1357                 mlog_errno(ret);
1358                 goto out;
1359         }
1360
1361         di = (struct ocfs2_dinode *)di_bh->b_data;
1362         data = &di->id2.i_data;
1363
1364         ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1365                                    i_size_read(dir));
1366
1367         brelse(di_bh);
1368 out:
1369         return ret;
1370 }
1371
1372 static inline int ocfs2_delete_entry_el(handle_t *handle,
1373                                         struct inode *dir,
1374                                         struct ocfs2_dir_entry *de_del,
1375                                         struct buffer_head *bh)
1376 {
1377         return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1378                                     bh->b_size);
1379 }
1380
1381 /*
1382  * Delete a directory entry. Hide the details of directory
1383  * implementation from the caller.
1384  */
1385 int ocfs2_delete_entry(handle_t *handle,
1386                        struct inode *dir,
1387                        struct ocfs2_dir_lookup_result *res)
1388 {
1389         if (ocfs2_dir_indexed(dir))
1390                 return ocfs2_delete_entry_dx(handle, dir, res);
1391
1392         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1393                 return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1394                                              res->dl_leaf_bh);
1395
1396         return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1397                                      res->dl_leaf_bh);
1398 }
1399
1400 /*
1401  * Check whether 'de' has enough room to hold an entry of
1402  * 'new_rec_len' bytes.
1403  */
1404 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1405                                          unsigned int new_rec_len)
1406 {
1407         unsigned int de_really_used;
1408
1409         /* Check whether this is an empty record with enough space */
1410         if (le64_to_cpu(de->inode) == 0 &&
1411             le16_to_cpu(de->rec_len) >= new_rec_len)
1412                 return 1;
1413
1414         /*
1415          * Record might have free space at the end which we can
1416          * use.
1417          */
1418         de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1419         if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1420             return 1;
1421
1422         return 0;
1423 }
1424
1425 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1426                                           struct ocfs2_dx_entry *dx_new_entry)
1427 {
1428         int i;
1429
1430         i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1431         dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1432
1433         le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1434 }
1435
1436 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1437                                        struct ocfs2_dx_hinfo *hinfo,
1438                                        u64 dirent_blk)
1439 {
1440         int i;
1441         struct ocfs2_dx_entry *dx_entry;
1442
1443         i = le16_to_cpu(entry_list->de_num_used);
1444         dx_entry = &entry_list->de_entries[i];
1445
1446         memset(dx_entry, 0, sizeof(*dx_entry));
1447         dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1448         dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1449         dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1450
1451         le16_add_cpu(&entry_list->de_num_used, 1);
1452 }
1453
1454 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1455                                       struct ocfs2_dx_hinfo *hinfo,
1456                                       u64 dirent_blk,
1457                                       struct buffer_head *dx_leaf_bh)
1458 {
1459         int ret;
1460         struct ocfs2_dx_leaf *dx_leaf;
1461
1462         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1463                                       OCFS2_JOURNAL_ACCESS_WRITE);
1464         if (ret) {
1465                 mlog_errno(ret);
1466                 goto out;
1467         }
1468
1469         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1470         ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1471         ocfs2_journal_dirty(handle, dx_leaf_bh);
1472
1473 out:
1474         return ret;
1475 }
1476
1477 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1478                                         struct ocfs2_dx_hinfo *hinfo,
1479                                         u64 dirent_blk,
1480                                         struct ocfs2_dx_root_block *dx_root)
1481 {
1482         ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1483 }
1484
1485 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1486                                struct ocfs2_dir_lookup_result *lookup)
1487 {
1488         int ret = 0;
1489         struct ocfs2_dx_root_block *dx_root;
1490         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1491
1492         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1493                                       OCFS2_JOURNAL_ACCESS_WRITE);
1494         if (ret) {
1495                 mlog_errno(ret);
1496                 goto out;
1497         }
1498
1499         dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1500         if (ocfs2_dx_root_inline(dx_root)) {
1501                 ocfs2_dx_inline_root_insert(dir, handle,
1502                                             &lookup->dl_hinfo,
1503                                             lookup->dl_leaf_bh->b_blocknr,
1504                                             dx_root);
1505         } else {
1506                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1507                                                  lookup->dl_leaf_bh->b_blocknr,
1508                                                  lookup->dl_dx_leaf_bh);
1509                 if (ret)
1510                         goto out;
1511         }
1512
1513         le32_add_cpu(&dx_root->dr_num_entries, 1);
1514         ocfs2_journal_dirty(handle, dx_root_bh);
1515
1516 out:
1517         return ret;
1518 }
1519
1520 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1521                                        handle_t *handle,
1522                                        struct ocfs2_dir_lookup_result *lookup)
1523 {
1524         struct ocfs2_dir_block_trailer *trailer, *prev;
1525         struct ocfs2_dx_root_block *dx_root;
1526         struct buffer_head *bh;
1527
1528         trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1529
1530         if (ocfs2_free_list_at_root(lookup)) {
1531                 bh = lookup->dl_dx_root_bh;
1532                 dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1533                 dx_root->dr_free_blk = trailer->db_free_next;
1534         } else {
1535                 bh = lookup->dl_prev_leaf_bh;
1536                 prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1537                 prev->db_free_next = trailer->db_free_next;
1538         }
1539
1540         trailer->db_free_rec_len = cpu_to_le16(0);
1541         trailer->db_free_next = cpu_to_le64(0);
1542
1543         ocfs2_journal_dirty(handle, bh);
1544         ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1545 }
1546
1547 /*
1548  * This expects that a journal write has been reserved on
1549  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1550  */
1551 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1552                                    struct ocfs2_dir_lookup_result *lookup)
1553 {
1554         int max_rec_len;
1555         struct ocfs2_dir_block_trailer *trailer;
1556
1557         /* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1558         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1559         if (max_rec_len) {
1560                 /*
1561                  * There's still room in this block, so no need to remove it
1562                  * from the free list. In this case, we just want to update
1563                  * the rec len accounting.
1564                  */
1565                 trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1566                 trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1567                 ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1568         } else {
1569                 ocfs2_remove_block_from_free_list(dir, handle, lookup);
1570         }
1571 }
1572
1573 /* we don't always have a dentry for what we want to add, so people
1574  * like orphan dir can call this instead.
1575  *
1576  * The lookup context must have been filled from
1577  * ocfs2_prepare_dir_for_insert.
1578  */
1579 int __ocfs2_add_entry(handle_t *handle,
1580                       struct inode *dir,
1581                       const char *name, int namelen,
1582                       struct inode *inode, u64 blkno,
1583                       struct buffer_head *parent_fe_bh,
1584                       struct ocfs2_dir_lookup_result *lookup)
1585 {
1586         unsigned long offset;
1587         unsigned short rec_len;
1588         struct ocfs2_dir_entry *de, *de1;
1589         struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1590         struct super_block *sb = dir->i_sb;
1591         int retval;
1592         unsigned int size = sb->s_blocksize;
1593         struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1594         char *data_start = insert_bh->b_data;
1595
1596         if (ocfs2_dir_indexed(dir)) {
1597                 struct buffer_head *bh;
1598
1599                 /*
1600                  * An indexed dir may require that we update the free space
1601                  * list. Reserve a write to the previous node in the list so
1602                  * that we don't fail later.
1603                  *
1604                  * XXX: This can be either a dx_root_block, or an unindexed
1605                  * directory tree leaf block.
1606                  */
1607                 if (ocfs2_free_list_at_root(lookup)) {
1608                         bh = lookup->dl_dx_root_bh;
1609                         retval = ocfs2_journal_access_dr(handle,
1610                                                  INODE_CACHE(dir), bh,
1611                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1612                 } else {
1613                         bh = lookup->dl_prev_leaf_bh;
1614                         retval = ocfs2_journal_access_db(handle,
1615                                                  INODE_CACHE(dir), bh,
1616                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1617                 }
1618                 if (retval) {
1619                         mlog_errno(retval);
1620                         return retval;
1621                 }
1622         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1623                 data_start = di->id2.i_data.id_data;
1624                 size = i_size_read(dir);
1625
1626                 BUG_ON(insert_bh != parent_fe_bh);
1627         }
1628
1629         rec_len = OCFS2_DIR_REC_LEN(namelen);
1630         offset = 0;
1631         de = (struct ocfs2_dir_entry *) data_start;
1632         while (1) {
1633                 BUG_ON((char *)de >= (size + data_start));
1634
1635                 /* These checks should've already been passed by the
1636                  * prepare function, but I guess we can leave them
1637                  * here anyway. */
1638                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1639                         retval = -ENOENT;
1640                         goto bail;
1641                 }
1642                 if (ocfs2_match(namelen, name, de)) {
1643                         retval = -EEXIST;
1644                         goto bail;
1645                 }
1646
1647                 /* We're guaranteed that we should have space, so we
1648                  * can't possibly have hit the trailer...right? */
1649                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1650                                 "Hit dir trailer trying to insert %.*s "
1651                                 "(namelen %d) into directory %llu.  "
1652                                 "offset is %lu, trailer offset is %d\n",
1653                                 namelen, name, namelen,
1654                                 (unsigned long long)parent_fe_bh->b_blocknr,
1655                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1656
1657                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1658                         inode_set_mtime_to_ts(dir,
1659                                               inode_set_ctime_current(dir));
1660                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1661                         if (retval < 0) {
1662                                 mlog_errno(retval);
1663                                 goto bail;
1664                         }
1665
1666                         if (insert_bh == parent_fe_bh)
1667                                 retval = ocfs2_journal_access_di(handle,
1668                                                                  INODE_CACHE(dir),
1669                                                                  insert_bh,
1670                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1671                         else {
1672                                 retval = ocfs2_journal_access_db(handle,
1673                                                                  INODE_CACHE(dir),
1674                                                                  insert_bh,
1675                                               OCFS2_JOURNAL_ACCESS_WRITE);
1676
1677                                 if (!retval && ocfs2_dir_indexed(dir))
1678                                         retval = ocfs2_dx_dir_insert(dir,
1679                                                                 handle,
1680                                                                 lookup);
1681                         }
1682
1683                         if (retval) {
1684                                 mlog_errno(retval);
1685                                 goto bail;
1686                         }
1687
1688                         /* By now the buffer is marked for journaling */
1689                         offset += le16_to_cpu(de->rec_len);
1690                         if (le64_to_cpu(de->inode)) {
1691                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1692                                         OCFS2_DIR_REC_LEN(de->name_len));
1693                                 de1->rec_len =
1694                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1695                                         OCFS2_DIR_REC_LEN(de->name_len));
1696                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1697                                 de = de1;
1698                         }
1699                         de->file_type = FT_UNKNOWN;
1700                         if (blkno) {
1701                                 de->inode = cpu_to_le64(blkno);
1702                                 ocfs2_set_de_type(de, inode->i_mode);
1703                         } else
1704                                 de->inode = 0;
1705                         de->name_len = namelen;
1706                         memcpy(de->name, name, namelen);
1707
1708                         if (ocfs2_dir_indexed(dir))
1709                                 ocfs2_recalc_free_list(dir, handle, lookup);
1710
1711                         inode_inc_iversion(dir);
1712                         ocfs2_journal_dirty(handle, insert_bh);
1713                         retval = 0;
1714                         goto bail;
1715                 }
1716
1717                 offset += le16_to_cpu(de->rec_len);
1718                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1719         }
1720
1721         /* when you think about it, the assert above should prevent us
1722          * from ever getting here. */
1723         retval = -ENOSPC;
1724 bail:
1725         if (retval)
1726                 mlog_errno(retval);
1727
1728         return retval;
1729 }
1730
1731 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1732                                     u64 *f_version,
1733                                     struct dir_context *ctx)
1734 {
1735         int ret, i;
1736         unsigned long offset = ctx->pos;
1737         struct buffer_head *di_bh = NULL;
1738         struct ocfs2_dinode *di;
1739         struct ocfs2_inline_data *data;
1740         struct ocfs2_dir_entry *de;
1741
1742         ret = ocfs2_read_inode_block(inode, &di_bh);
1743         if (ret) {
1744                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1745                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1746                 goto out;
1747         }
1748
1749         di = (struct ocfs2_dinode *)di_bh->b_data;
1750         data = &di->id2.i_data;
1751
1752         while (ctx->pos < i_size_read(inode)) {
1753                 /* If the dir block has changed since the last call to
1754                  * readdir(2), then we might be pointing to an invalid
1755                  * dirent right now.  Scan from the start of the block
1756                  * to make sure. */
1757                 if (!inode_eq_iversion(inode, *f_version)) {
1758                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1759                                 de = (struct ocfs2_dir_entry *)
1760                                         (data->id_data + i);
1761                                 /* It's too expensive to do a full
1762                                  * dirent test each time round this
1763                                  * loop, but we do have to test at
1764                                  * least that it is non-zero.  A
1765                                  * failure will be detected in the
1766                                  * dirent test below. */
1767                                 if (le16_to_cpu(de->rec_len) <
1768                                     OCFS2_DIR_REC_LEN(1))
1769                                         break;
1770                                 i += le16_to_cpu(de->rec_len);
1771                         }
1772                         ctx->pos = offset = i;
1773                         *f_version = inode_query_iversion(inode);
1774                 }
1775
1776                 de = (struct ocfs2_dir_entry *) (data->id_data + ctx->pos);
1777                 if (!ocfs2_check_dir_entry(inode, de, di_bh, ctx->pos)) {
1778                         /* On error, skip the f_pos to the end. */
1779                         ctx->pos = i_size_read(inode);
1780                         break;
1781                 }
1782                 offset += le16_to_cpu(de->rec_len);
1783                 if (le64_to_cpu(de->inode)) {
1784                         if (!dir_emit(ctx, de->name, de->name_len,
1785                                       le64_to_cpu(de->inode),
1786                                       fs_ftype_to_dtype(de->file_type)))
1787                                 goto out;
1788                 }
1789                 ctx->pos += le16_to_cpu(de->rec_len);
1790         }
1791 out:
1792         brelse(di_bh);
1793         return 0;
1794 }
1795
1796 /*
1797  * NOTE: This function can be called against unindexed directories,
1798  * and indexed ones.
1799  */
1800 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1801                                     u64 *f_version,
1802                                     struct dir_context *ctx,
1803                                     bool persist)
1804 {
1805         unsigned long offset, blk, last_ra_blk = 0;
1806         int i;
1807         struct buffer_head * bh, * tmp;
1808         struct ocfs2_dir_entry * de;
1809         struct super_block * sb = inode->i_sb;
1810         unsigned int ra_sectors = 16;
1811         int stored = 0;
1812
1813         bh = NULL;
1814
1815         offset = ctx->pos & (sb->s_blocksize - 1);
1816
1817         while (ctx->pos < i_size_read(inode)) {
1818                 blk = ctx->pos >> sb->s_blocksize_bits;
1819                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1820                         /* Skip the corrupt dirblock and keep trying */
1821                         ctx->pos += sb->s_blocksize - offset;
1822                         continue;
1823                 }
1824
1825                 /* The idea here is to begin with 8k read-ahead and to stay
1826                  * 4k ahead of our current position.
1827                  *
1828                  * TODO: Use the pagecache for this. We just need to
1829                  * make sure it's cluster-safe... */
1830                 if (!last_ra_blk
1831                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1832                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1833                              i > 0; i--) {
1834                                 tmp = NULL;
1835                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1836                                                           OCFS2_BH_READAHEAD))
1837                                         brelse(tmp);
1838                         }
1839                         last_ra_blk = blk;
1840                         ra_sectors = 8;
1841                 }
1842
1843                 /* If the dir block has changed since the last call to
1844                  * readdir(2), then we might be pointing to an invalid
1845                  * dirent right now.  Scan from the start of the block
1846                  * to make sure. */
1847                 if (!inode_eq_iversion(inode, *f_version)) {
1848                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1849                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1850                                 /* It's too expensive to do a full
1851                                  * dirent test each time round this
1852                                  * loop, but we do have to test at
1853                                  * least that it is non-zero.  A
1854                                  * failure will be detected in the
1855                                  * dirent test below. */
1856                                 if (le16_to_cpu(de->rec_len) <
1857                                     OCFS2_DIR_REC_LEN(1))
1858                                         break;
1859                                 i += le16_to_cpu(de->rec_len);
1860                         }
1861                         offset = i;
1862                         ctx->pos = (ctx->pos & ~(sb->s_blocksize - 1))
1863                                 | offset;
1864                         *f_version = inode_query_iversion(inode);
1865                 }
1866
1867                 while (ctx->pos < i_size_read(inode)
1868                        && offset < sb->s_blocksize) {
1869                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1870                         if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1871                                 /* On error, skip the f_pos to the
1872                                    next block. */
1873                                 ctx->pos = (ctx->pos | (sb->s_blocksize - 1)) + 1;
1874                                 break;
1875                         }
1876                         if (le64_to_cpu(de->inode)) {
1877                                 if (!dir_emit(ctx, de->name,
1878                                                 de->name_len,
1879                                                 le64_to_cpu(de->inode),
1880                                         fs_ftype_to_dtype(de->file_type))) {
1881                                         brelse(bh);
1882                                         return 0;
1883                                 }
1884                                 stored++;
1885                         }
1886                         offset += le16_to_cpu(de->rec_len);
1887                         ctx->pos += le16_to_cpu(de->rec_len);
1888                 }
1889                 offset = 0;
1890                 brelse(bh);
1891                 bh = NULL;
1892                 if (!persist && stored)
1893                         break;
1894         }
1895         return 0;
1896 }
1897
1898 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1899                                  struct dir_context *ctx,
1900                                  bool persist)
1901 {
1902         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1903                 return ocfs2_dir_foreach_blk_id(inode, f_version, ctx);
1904         return ocfs2_dir_foreach_blk_el(inode, f_version, ctx, persist);
1905 }
1906
1907 /*
1908  * This is intended to be called from inside other kernel functions,
1909  * so we fake some arguments.
1910  */
1911 int ocfs2_dir_foreach(struct inode *inode, struct dir_context *ctx)
1912 {
1913         u64 version = inode_query_iversion(inode);
1914         ocfs2_dir_foreach_blk(inode, &version, ctx, true);
1915         return 0;
1916 }
1917
1918 /*
1919  * ocfs2_readdir()
1920  *
1921  */
1922 int ocfs2_readdir(struct file *file, struct dir_context *ctx)
1923 {
1924         int error = 0;
1925         struct inode *inode = file_inode(file);
1926         int lock_level = 0;
1927
1928         trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
1929
1930         error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level, 1);
1931         if (lock_level && error >= 0) {
1932                 /* We release EX lock which used to update atime
1933                  * and get PR lock again to reduce contention
1934                  * on commonly accessed directories. */
1935                 ocfs2_inode_unlock(inode, 1);
1936                 lock_level = 0;
1937                 error = ocfs2_inode_lock(inode, NULL, 0);
1938         }
1939         if (error < 0) {
1940                 if (error != -ENOENT)
1941                         mlog_errno(error);
1942                 /* we haven't got any yet, so propagate the error. */
1943                 goto bail_nolock;
1944         }
1945
1946         error = ocfs2_dir_foreach_blk(inode, &file->f_version, ctx, false);
1947
1948         ocfs2_inode_unlock(inode, lock_level);
1949         if (error)
1950                 mlog_errno(error);
1951
1952 bail_nolock:
1953
1954         return error;
1955 }
1956
1957 /*
1958  * NOTE: this should always be called with parent dir i_rwsem taken.
1959  */
1960 int ocfs2_find_files_on_disk(const char *name,
1961                              int namelen,
1962                              u64 *blkno,
1963                              struct inode *inode,
1964                              struct ocfs2_dir_lookup_result *lookup)
1965 {
1966         int status = -ENOENT;
1967
1968         trace_ocfs2_find_files_on_disk(namelen, name, blkno,
1969                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1970
1971         status = ocfs2_find_entry(name, namelen, inode, lookup);
1972         if (status)
1973                 goto leave;
1974
1975         *blkno = le64_to_cpu(lookup->dl_entry->inode);
1976
1977         status = 0;
1978 leave:
1979
1980         return status;
1981 }
1982
1983 /*
1984  * Convenience function for callers which just want the block number
1985  * mapped to a name and don't require the full dirent info, etc.
1986  */
1987 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
1988                                int namelen, u64 *blkno)
1989 {
1990         int ret;
1991         struct ocfs2_dir_lookup_result lookup = { NULL, };
1992
1993         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
1994         ocfs2_free_dir_lookup_result(&lookup);
1995
1996         return ret;
1997 }
1998
1999 /* Check for a name within a directory.
2000  *
2001  * Return 0 if the name does not exist
2002  * Return -EEXIST if the directory contains the name
2003  *
2004  * Callers should have i_rwsem + a cluster lock on dir
2005  */
2006 int ocfs2_check_dir_for_entry(struct inode *dir,
2007                               const char *name,
2008                               int namelen)
2009 {
2010         int ret = 0;
2011         struct ocfs2_dir_lookup_result lookup = { NULL, };
2012
2013         trace_ocfs2_check_dir_for_entry(
2014                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2015
2016         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0) {
2017                 ret = -EEXIST;
2018                 mlog_errno(ret);
2019         }
2020
2021         ocfs2_free_dir_lookup_result(&lookup);
2022
2023         return ret;
2024 }
2025
2026 struct ocfs2_empty_dir_priv {
2027         struct dir_context ctx;
2028         unsigned seen_dot;
2029         unsigned seen_dot_dot;
2030         unsigned seen_other;
2031         unsigned dx_dir;
2032 };
2033 static bool ocfs2_empty_dir_filldir(struct dir_context *ctx, const char *name,
2034                                    int name_len, loff_t pos, u64 ino,
2035                                    unsigned type)
2036 {
2037         struct ocfs2_empty_dir_priv *p =
2038                 container_of(ctx, struct ocfs2_empty_dir_priv, ctx);
2039
2040         /*
2041          * Check the positions of "." and ".." records to be sure
2042          * they're in the correct place.
2043          *
2044          * Indexed directories don't need to proceed past the first
2045          * two entries, so we end the scan after seeing '..'. Despite
2046          * that, we allow the scan to proceed In the event that we
2047          * have a corrupted indexed directory (no dot or dot dot
2048          * entries). This allows us to double check for existing
2049          * entries which might not have been found in the index.
2050          */
2051         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2052                 p->seen_dot = 1;
2053                 return true;
2054         }
2055
2056         if (name_len == 2 && !strncmp("..", name, 2) &&
2057             pos == OCFS2_DIR_REC_LEN(1)) {
2058                 p->seen_dot_dot = 1;
2059
2060                 if (p->dx_dir && p->seen_dot)
2061                         return false;
2062
2063                 return true;
2064         }
2065
2066         p->seen_other = 1;
2067         return false;
2068 }
2069
2070 static int ocfs2_empty_dir_dx(struct inode *inode,
2071                               struct ocfs2_empty_dir_priv *priv)
2072 {
2073         int ret;
2074         struct buffer_head *di_bh = NULL;
2075         struct buffer_head *dx_root_bh = NULL;
2076         struct ocfs2_dinode *di;
2077         struct ocfs2_dx_root_block *dx_root;
2078
2079         priv->dx_dir = 1;
2080
2081         ret = ocfs2_read_inode_block(inode, &di_bh);
2082         if (ret) {
2083                 mlog_errno(ret);
2084                 goto out;
2085         }
2086         di = (struct ocfs2_dinode *)di_bh->b_data;
2087
2088         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2089         if (ret) {
2090                 mlog_errno(ret);
2091                 goto out;
2092         }
2093         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2094
2095         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2096                 priv->seen_other = 1;
2097
2098 out:
2099         brelse(di_bh);
2100         brelse(dx_root_bh);
2101         return ret;
2102 }
2103
2104 /*
2105  * routine to check that the specified directory is empty (for rmdir)
2106  *
2107  * Returns 1 if dir is empty, zero otherwise.
2108  *
2109  * XXX: This is a performance problem for unindexed directories.
2110  */
2111 int ocfs2_empty_dir(struct inode *inode)
2112 {
2113         int ret;
2114         struct ocfs2_empty_dir_priv priv = {
2115                 .ctx.actor = ocfs2_empty_dir_filldir,
2116         };
2117
2118         if (ocfs2_dir_indexed(inode)) {
2119                 ret = ocfs2_empty_dir_dx(inode, &priv);
2120                 if (ret)
2121                         mlog_errno(ret);
2122                 /*
2123                  * We still run ocfs2_dir_foreach to get the checks
2124                  * for "." and "..".
2125                  */
2126         }
2127
2128         ret = ocfs2_dir_foreach(inode, &priv.ctx);
2129         if (ret)
2130                 mlog_errno(ret);
2131
2132         if (!priv.seen_dot || !priv.seen_dot_dot) {
2133                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2134                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2135                 /*
2136                  * XXX: Is it really safe to allow an unlink to continue?
2137                  */
2138                 return 1;
2139         }
2140
2141         return !priv.seen_other;
2142 }
2143
2144 /*
2145  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2146  * "..", which might be used during creation of a directory with a trailing
2147  * header. It is otherwise safe to ignore the return code.
2148  */
2149 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2150                                                           struct inode *parent,
2151                                                           char *start,
2152                                                           unsigned int size)
2153 {
2154         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2155
2156         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2157         de->name_len = 1;
2158         de->rec_len =
2159                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2160         strcpy(de->name, ".");
2161         ocfs2_set_de_type(de, S_IFDIR);
2162
2163         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2164         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2165         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2166         de->name_len = 2;
2167         strcpy(de->name, "..");
2168         ocfs2_set_de_type(de, S_IFDIR);
2169
2170         return de;
2171 }
2172
2173 /*
2174  * This works together with code in ocfs2_mknod_locked() which sets
2175  * the inline-data flag and initializes the inline-data section.
2176  */
2177 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2178                                  handle_t *handle,
2179                                  struct inode *parent,
2180                                  struct inode *inode,
2181                                  struct buffer_head *di_bh)
2182 {
2183         int ret;
2184         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2185         struct ocfs2_inline_data *data = &di->id2.i_data;
2186         unsigned int size = le16_to_cpu(data->id_count);
2187
2188         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2189                                       OCFS2_JOURNAL_ACCESS_WRITE);
2190         if (ret) {
2191                 mlog_errno(ret);
2192                 goto out;
2193         }
2194
2195         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2196         ocfs2_journal_dirty(handle, di_bh);
2197
2198         i_size_write(inode, size);
2199         set_nlink(inode, 2);
2200         inode->i_blocks = ocfs2_inode_sector_count(inode);
2201
2202         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2203         if (ret < 0)
2204                 mlog_errno(ret);
2205
2206 out:
2207         return ret;
2208 }
2209
2210 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2211                                  handle_t *handle,
2212                                  struct inode *parent,
2213                                  struct inode *inode,
2214                                  struct buffer_head *fe_bh,
2215                                  struct ocfs2_alloc_context *data_ac,
2216                                  struct buffer_head **ret_new_bh)
2217 {
2218         int status;
2219         unsigned int size = osb->sb->s_blocksize;
2220         struct buffer_head *new_bh = NULL;
2221         struct ocfs2_dir_entry *de;
2222
2223         if (ocfs2_new_dir_wants_trailer(inode))
2224                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2225
2226         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2227                                      data_ac, NULL, &new_bh);
2228         if (status < 0) {
2229                 mlog_errno(status);
2230                 goto bail;
2231         }
2232
2233         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2234
2235         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2236                                          OCFS2_JOURNAL_ACCESS_CREATE);
2237         if (status < 0) {
2238                 mlog_errno(status);
2239                 goto bail;
2240         }
2241         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2242
2243         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2244         if (ocfs2_new_dir_wants_trailer(inode)) {
2245                 int size = le16_to_cpu(de->rec_len);
2246
2247                 /*
2248                  * Figure out the size of the hole left over after
2249                  * insertion of '.' and '..'. The trailer wants this
2250                  * information.
2251                  */
2252                 size -= OCFS2_DIR_REC_LEN(2);
2253                 size -= sizeof(struct ocfs2_dir_block_trailer);
2254
2255                 ocfs2_init_dir_trailer(inode, new_bh, size);
2256         }
2257
2258         ocfs2_journal_dirty(handle, new_bh);
2259
2260         i_size_write(inode, inode->i_sb->s_blocksize);
2261         set_nlink(inode, 2);
2262         inode->i_blocks = ocfs2_inode_sector_count(inode);
2263         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2264         if (status < 0) {
2265                 mlog_errno(status);
2266                 goto bail;
2267         }
2268
2269         status = 0;
2270         if (ret_new_bh) {
2271                 *ret_new_bh = new_bh;
2272                 new_bh = NULL;
2273         }
2274 bail:
2275         brelse(new_bh);
2276
2277         return status;
2278 }
2279
2280 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2281                                      handle_t *handle, struct inode *dir,
2282                                      struct buffer_head *di_bh,
2283                                      struct buffer_head *dirdata_bh,
2284                                      struct ocfs2_alloc_context *meta_ac,
2285                                      int dx_inline, u32 num_entries,
2286                                      struct buffer_head **ret_dx_root_bh)
2287 {
2288         int ret;
2289         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2290         u16 dr_suballoc_bit;
2291         u64 suballoc_loc, dr_blkno;
2292         unsigned int num_bits;
2293         struct buffer_head *dx_root_bh = NULL;
2294         struct ocfs2_dx_root_block *dx_root;
2295         struct ocfs2_dir_block_trailer *trailer =
2296                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2297
2298         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2299                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2300         if (ret) {
2301                 mlog_errno(ret);
2302                 goto out;
2303         }
2304
2305         trace_ocfs2_dx_dir_attach_index(
2306                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2307                                 (unsigned long long)dr_blkno);
2308
2309         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2310         if (dx_root_bh == NULL) {
2311                 ret = -ENOMEM;
2312                 goto out;
2313         }
2314         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2315
2316         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2317                                       OCFS2_JOURNAL_ACCESS_CREATE);
2318         if (ret < 0) {
2319                 mlog_errno(ret);
2320                 goto out;
2321         }
2322
2323         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2324         memset(dx_root, 0, osb->sb->s_blocksize);
2325         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2326         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2327         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2328         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2329         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2330         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2331         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2332         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2333         if (le16_to_cpu(trailer->db_free_rec_len))
2334                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2335         else
2336                 dx_root->dr_free_blk = cpu_to_le64(0);
2337
2338         if (dx_inline) {
2339                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2340                 dx_root->dr_entries.de_count =
2341                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2342         } else {
2343                 dx_root->dr_list.l_count =
2344                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2345         }
2346         ocfs2_journal_dirty(handle, dx_root_bh);
2347
2348         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2349                                       OCFS2_JOURNAL_ACCESS_CREATE);
2350         if (ret) {
2351                 mlog_errno(ret);
2352                 goto out;
2353         }
2354
2355         di->i_dx_root = cpu_to_le64(dr_blkno);
2356
2357         spin_lock(&OCFS2_I(dir)->ip_lock);
2358         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2359         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2360         spin_unlock(&OCFS2_I(dir)->ip_lock);
2361
2362         ocfs2_journal_dirty(handle, di_bh);
2363
2364         *ret_dx_root_bh = dx_root_bh;
2365         dx_root_bh = NULL;
2366
2367 out:
2368         brelse(dx_root_bh);
2369         return ret;
2370 }
2371
2372 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2373                                        handle_t *handle, struct inode *dir,
2374                                        struct buffer_head **dx_leaves,
2375                                        int num_dx_leaves, u64 start_blk)
2376 {
2377         int ret, i;
2378         struct ocfs2_dx_leaf *dx_leaf;
2379         struct buffer_head *bh;
2380
2381         for (i = 0; i < num_dx_leaves; i++) {
2382                 bh = sb_getblk(osb->sb, start_blk + i);
2383                 if (bh == NULL) {
2384                         ret = -ENOMEM;
2385                         goto out;
2386                 }
2387                 dx_leaves[i] = bh;
2388
2389                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2390
2391                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2392                                               OCFS2_JOURNAL_ACCESS_CREATE);
2393                 if (ret < 0) {
2394                         mlog_errno(ret);
2395                         goto out;
2396                 }
2397
2398                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2399
2400                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2401                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2402                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2403                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2404                 dx_leaf->dl_list.de_count =
2405                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2406
2407                 trace_ocfs2_dx_dir_format_cluster(
2408                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2409                                 (unsigned long long)bh->b_blocknr,
2410                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2411
2412                 ocfs2_journal_dirty(handle, bh);
2413         }
2414
2415         ret = 0;
2416 out:
2417         return ret;
2418 }
2419
2420 /*
2421  * Allocates and formats a new cluster for use in an indexed dir
2422  * leaf. This version will not do the extent insert, so that it can be
2423  * used by operations which need careful ordering.
2424  */
2425 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2426                                       u32 cpos, handle_t *handle,
2427                                       struct ocfs2_alloc_context *data_ac,
2428                                       struct buffer_head **dx_leaves,
2429                                       int num_dx_leaves, u64 *ret_phys_blkno)
2430 {
2431         int ret;
2432         u32 phys, num;
2433         u64 phys_blkno;
2434         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2435
2436         /*
2437          * XXX: For create, this should claim cluster for the index
2438          * *before* the unindexed insert so that we have a better
2439          * chance of contiguousness as the directory grows in number
2440          * of entries.
2441          */
2442         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2443         if (ret) {
2444                 mlog_errno(ret);
2445                 goto out;
2446         }
2447
2448         /*
2449          * Format the new cluster first. That way, we're inserting
2450          * valid data.
2451          */
2452         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2453         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2454                                           num_dx_leaves, phys_blkno);
2455         if (ret) {
2456                 mlog_errno(ret);
2457                 goto out;
2458         }
2459
2460         *ret_phys_blkno = phys_blkno;
2461 out:
2462         return ret;
2463 }
2464
2465 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2466                                     struct ocfs2_extent_tree *et,
2467                                     u32 cpos, handle_t *handle,
2468                                     struct ocfs2_alloc_context *data_ac,
2469                                     struct ocfs2_alloc_context *meta_ac,
2470                                     struct buffer_head **dx_leaves,
2471                                     int num_dx_leaves)
2472 {
2473         int ret;
2474         u64 phys_blkno;
2475
2476         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2477                                          num_dx_leaves, &phys_blkno);
2478         if (ret) {
2479                 mlog_errno(ret);
2480                 goto out;
2481         }
2482
2483         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2484                                   meta_ac);
2485         if (ret)
2486                 mlog_errno(ret);
2487 out:
2488         return ret;
2489 }
2490
2491 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2492                                                         int *ret_num_leaves)
2493 {
2494         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2495         struct buffer_head **dx_leaves;
2496
2497         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2498                             GFP_NOFS);
2499         if (dx_leaves && ret_num_leaves)
2500                 *ret_num_leaves = num_dx_leaves;
2501
2502         return dx_leaves;
2503 }
2504
2505 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2506                                  handle_t *handle,
2507                                  struct inode *parent,
2508                                  struct inode *inode,
2509                                  struct buffer_head *di_bh,
2510                                  struct ocfs2_alloc_context *data_ac,
2511                                  struct ocfs2_alloc_context *meta_ac)
2512 {
2513         int ret;
2514         struct buffer_head *leaf_bh = NULL;
2515         struct buffer_head *dx_root_bh = NULL;
2516         struct ocfs2_dx_hinfo hinfo;
2517         struct ocfs2_dx_root_block *dx_root;
2518         struct ocfs2_dx_entry_list *entry_list;
2519
2520         /*
2521          * Our strategy is to create the directory as though it were
2522          * unindexed, then add the index block. This works with very
2523          * little complication since the state of a new directory is a
2524          * very well known quantity.
2525          *
2526          * Essentially, we have two dirents ("." and ".."), in the 1st
2527          * block which need indexing. These are easily inserted into
2528          * the index block.
2529          */
2530
2531         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2532                                     data_ac, &leaf_bh);
2533         if (ret) {
2534                 mlog_errno(ret);
2535                 goto out;
2536         }
2537
2538         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2539                                         meta_ac, 1, 2, &dx_root_bh);
2540         if (ret) {
2541                 mlog_errno(ret);
2542                 goto out;
2543         }
2544         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2545         entry_list = &dx_root->dr_entries;
2546
2547         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2548         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2549         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2550
2551         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2552         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2553
2554 out:
2555         brelse(dx_root_bh);
2556         brelse(leaf_bh);
2557         return ret;
2558 }
2559
2560 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2561                        handle_t *handle,
2562                        struct inode *parent,
2563                        struct inode *inode,
2564                        struct buffer_head *fe_bh,
2565                        struct ocfs2_alloc_context *data_ac,
2566                        struct ocfs2_alloc_context *meta_ac)
2567
2568 {
2569         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2570
2571         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2572                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2573
2574         if (ocfs2_supports_indexed_dirs(osb))
2575                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2576                                              data_ac, meta_ac);
2577
2578         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2579                                      data_ac, NULL);
2580 }
2581
2582 static int ocfs2_dx_dir_index_block(struct inode *dir,
2583                                     handle_t *handle,
2584                                     struct buffer_head **dx_leaves,
2585                                     int num_dx_leaves,
2586                                     u32 *num_dx_entries,
2587                                     struct buffer_head *dirent_bh)
2588 {
2589         int ret = 0, namelen, i;
2590         char *de_buf, *limit;
2591         struct ocfs2_dir_entry *de;
2592         struct buffer_head *dx_leaf_bh;
2593         struct ocfs2_dx_hinfo hinfo;
2594         u64 dirent_blk = dirent_bh->b_blocknr;
2595
2596         de_buf = dirent_bh->b_data;
2597         limit = de_buf + dir->i_sb->s_blocksize;
2598
2599         while (de_buf < limit) {
2600                 de = (struct ocfs2_dir_entry *)de_buf;
2601
2602                 namelen = de->name_len;
2603                 if (!namelen || !de->inode)
2604                         goto inc;
2605
2606                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2607
2608                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2609                 dx_leaf_bh = dx_leaves[i];
2610
2611                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2612                                                  dirent_blk, dx_leaf_bh);
2613                 if (ret) {
2614                         mlog_errno(ret);
2615                         goto out;
2616                 }
2617
2618                 *num_dx_entries = *num_dx_entries + 1;
2619
2620 inc:
2621                 de_buf += le16_to_cpu(de->rec_len);
2622         }
2623
2624 out:
2625         return ret;
2626 }
2627
2628 /*
2629  * XXX: This expects dx_root_bh to already be part of the transaction.
2630  */
2631 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2632                                          struct buffer_head *dx_root_bh,
2633                                          struct buffer_head *dirent_bh)
2634 {
2635         char *de_buf, *limit;
2636         struct ocfs2_dx_root_block *dx_root;
2637         struct ocfs2_dir_entry *de;
2638         struct ocfs2_dx_hinfo hinfo;
2639         u64 dirent_blk = dirent_bh->b_blocknr;
2640
2641         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2642
2643         de_buf = dirent_bh->b_data;
2644         limit = de_buf + dir->i_sb->s_blocksize;
2645
2646         while (de_buf < limit) {
2647                 de = (struct ocfs2_dir_entry *)de_buf;
2648
2649                 if (!de->name_len || !de->inode)
2650                         goto inc;
2651
2652                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2653
2654                 trace_ocfs2_dx_dir_index_root_block(
2655                                 (unsigned long long)dir->i_ino,
2656                                 hinfo.major_hash, hinfo.minor_hash,
2657                                 de->name_len, de->name,
2658                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2659
2660                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2661                                            dirent_blk);
2662
2663                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2664 inc:
2665                 de_buf += le16_to_cpu(de->rec_len);
2666         }
2667 }
2668
2669 /*
2670  * Count the number of inline directory entries in di_bh and compare
2671  * them against the number of entries we can hold in an inline dx root
2672  * block.
2673  */
2674 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2675                                          struct buffer_head *di_bh)
2676 {
2677         int dirent_count = 0;
2678         char *de_buf, *limit;
2679         struct ocfs2_dir_entry *de;
2680         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2681
2682         de_buf = di->id2.i_data.id_data;
2683         limit = de_buf + i_size_read(dir);
2684
2685         while (de_buf < limit) {
2686                 de = (struct ocfs2_dir_entry *)de_buf;
2687
2688                 if (de->name_len && de->inode)
2689                         dirent_count++;
2690
2691                 de_buf += le16_to_cpu(de->rec_len);
2692         }
2693
2694         /* We are careful to leave room for one extra record. */
2695         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2696 }
2697
2698 /*
2699  * Expand rec_len of the rightmost dirent in a directory block so that it
2700  * contains the end of our valid space for dirents. We do this during
2701  * expansion from an inline directory to one with extents. The first dir block
2702  * in that case is taken from the inline data portion of the inode block.
2703  *
2704  * This will also return the largest amount of contiguous space for a dirent
2705  * in the block. That value is *not* necessarily the last dirent, even after
2706  * expansion. The directory indexing code wants this value for free space
2707  * accounting. We do this here since we're already walking the entire dir
2708  * block.
2709  *
2710  * We add the dir trailer if this filesystem wants it.
2711  */
2712 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2713                                              struct inode *dir)
2714 {
2715         struct super_block *sb = dir->i_sb;
2716         struct ocfs2_dir_entry *de;
2717         struct ocfs2_dir_entry *prev_de;
2718         char *de_buf, *limit;
2719         unsigned int new_size = sb->s_blocksize;
2720         unsigned int bytes, this_hole;
2721         unsigned int largest_hole = 0;
2722
2723         if (ocfs2_new_dir_wants_trailer(dir))
2724                 new_size = ocfs2_dir_trailer_blk_off(sb);
2725
2726         bytes = new_size - old_size;
2727
2728         limit = start + old_size;
2729         de_buf = start;
2730         de = (struct ocfs2_dir_entry *)de_buf;
2731         do {
2732                 this_hole = ocfs2_figure_dirent_hole(de);
2733                 if (this_hole > largest_hole)
2734                         largest_hole = this_hole;
2735
2736                 prev_de = de;
2737                 de_buf += le16_to_cpu(de->rec_len);
2738                 de = (struct ocfs2_dir_entry *)de_buf;
2739         } while (de_buf < limit);
2740
2741         le16_add_cpu(&prev_de->rec_len, bytes);
2742
2743         /* We need to double check this after modification of the final
2744          * dirent. */
2745         this_hole = ocfs2_figure_dirent_hole(prev_de);
2746         if (this_hole > largest_hole)
2747                 largest_hole = this_hole;
2748
2749         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2750                 return largest_hole;
2751         return 0;
2752 }
2753
2754 /*
2755  * We allocate enough clusters to fulfill "blocks_wanted", but set
2756  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2757  * rest automatically for us.
2758  *
2759  * *first_block_bh is a pointer to the 1st data block allocated to the
2760  *  directory.
2761  */
2762 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2763                                    unsigned int blocks_wanted,
2764                                    struct ocfs2_dir_lookup_result *lookup,
2765                                    struct buffer_head **first_block_bh)
2766 {
2767         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2768         struct super_block *sb = dir->i_sb;
2769         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2770                 credits = ocfs2_inline_to_extents_credits(sb);
2771         u64 dx_insert_blkno, blkno,
2772                 bytes = blocks_wanted << sb->s_blocksize_bits;
2773         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2774         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2775         struct ocfs2_alloc_context *data_ac = NULL;
2776         struct ocfs2_alloc_context *meta_ac = NULL;
2777         struct buffer_head *dirdata_bh = NULL;
2778         struct buffer_head *dx_root_bh = NULL;
2779         struct buffer_head **dx_leaves = NULL;
2780         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2781         handle_t *handle;
2782         struct ocfs2_extent_tree et;
2783         struct ocfs2_extent_tree dx_et;
2784         int did_quota = 0, bytes_allocated = 0;
2785
2786         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2787
2788         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2789         dx_alloc = 0;
2790
2791         down_write(&oi->ip_alloc_sem);
2792
2793         if (ocfs2_supports_indexed_dirs(osb)) {
2794                 credits += ocfs2_add_dir_index_credits(sb);
2795
2796                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2797                 if (!dx_inline) {
2798                         /* Add one more cluster for an index leaf */
2799                         dx_alloc++;
2800                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2801                                                                 &num_dx_leaves);
2802                         if (!dx_leaves) {
2803                                 ret = -ENOMEM;
2804                                 mlog_errno(ret);
2805                                 goto out;
2806                         }
2807                 }
2808
2809                 /* This gets us the dx_root */
2810                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2811                 if (ret) {
2812                         mlog_errno(ret);
2813                         goto out;
2814                 }
2815         }
2816
2817         /*
2818          * We should never need more than 2 clusters for the unindexed
2819          * tree - maximum dirent size is far less than one block. In
2820          * fact, the only time we'd need more than one cluster is if
2821          * blocksize == clustersize and the dirent won't fit in the
2822          * extra space that the expansion to a single block gives. As
2823          * of today, that only happens on 4k/4k file systems.
2824          */
2825         BUG_ON(alloc > 2);
2826
2827         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2828         if (ret) {
2829                 mlog_errno(ret);
2830                 goto out;
2831         }
2832
2833         /*
2834          * Prepare for worst case allocation scenario of two separate
2835          * extents in the unindexed tree.
2836          */
2837         if (alloc == 2)
2838                 credits += OCFS2_SUBALLOC_ALLOC;
2839
2840         handle = ocfs2_start_trans(osb, credits);
2841         if (IS_ERR(handle)) {
2842                 ret = PTR_ERR(handle);
2843                 mlog_errno(ret);
2844                 goto out;
2845         }
2846
2847         ret = dquot_alloc_space_nodirty(dir,
2848                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2849         if (ret)
2850                 goto out_commit;
2851         did_quota = 1;
2852
2853         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2854                 /*
2855                  * Allocate our index cluster first, to maximize the
2856                  * possibility that unindexed leaves grow
2857                  * contiguously.
2858                  */
2859                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2860                                                  dx_leaves, num_dx_leaves,
2861                                                  &dx_insert_blkno);
2862                 if (ret) {
2863                         mlog_errno(ret);
2864                         goto out_commit;
2865                 }
2866                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2867         }
2868
2869         /*
2870          * Try to claim as many clusters as the bitmap can give though
2871          * if we only get one now, that's enough to continue. The rest
2872          * will be claimed after the conversion to extents.
2873          */
2874         if (ocfs2_dir_resv_allowed(osb))
2875                 data_ac->ac_resv = &oi->ip_la_data_resv;
2876         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2877         if (ret) {
2878                 mlog_errno(ret);
2879                 goto out_commit;
2880         }
2881         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2882
2883         /*
2884          * Operations are carefully ordered so that we set up the new
2885          * data block first. The conversion from inline data to
2886          * extents follows.
2887          */
2888         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2889         dirdata_bh = sb_getblk(sb, blkno);
2890         if (!dirdata_bh) {
2891                 ret = -ENOMEM;
2892                 mlog_errno(ret);
2893                 goto out_commit;
2894         }
2895
2896         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2897
2898         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2899                                       OCFS2_JOURNAL_ACCESS_CREATE);
2900         if (ret) {
2901                 mlog_errno(ret);
2902                 goto out_commit;
2903         }
2904
2905         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2906         memset(dirdata_bh->b_data + i_size_read(dir), 0,
2907                sb->s_blocksize - i_size_read(dir));
2908         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2909         if (ocfs2_new_dir_wants_trailer(dir)) {
2910                 /*
2911                  * Prepare the dir trailer up front. It will otherwise look
2912                  * like a valid dirent. Even if inserting the index fails
2913                  * (unlikely), then all we'll have done is given first dir
2914                  * block a small amount of fragmentation.
2915                  */
2916                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2917         }
2918
2919         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2920         ocfs2_journal_dirty(handle, dirdata_bh);
2921
2922         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2923                 /*
2924                  * Dx dirs with an external cluster need to do this up
2925                  * front. Inline dx root's get handled later, after
2926                  * we've allocated our root block. We get passed back
2927                  * a total number of items so that dr_num_entries can
2928                  * be correctly set once the dx_root has been
2929                  * allocated.
2930                  */
2931                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2932                                                num_dx_leaves, &num_dx_entries,
2933                                                dirdata_bh);
2934                 if (ret) {
2935                         mlog_errno(ret);
2936                         goto out_commit;
2937                 }
2938         }
2939
2940         /*
2941          * Set extent, i_size, etc on the directory. After this, the
2942          * inode should contain the same exact dirents as before and
2943          * be fully accessible from system calls.
2944          *
2945          * We let the later dirent insert modify c/mtime - to the user
2946          * the data hasn't changed.
2947          */
2948         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2949                                       OCFS2_JOURNAL_ACCESS_CREATE);
2950         if (ret) {
2951                 mlog_errno(ret);
2952                 goto out_commit;
2953         }
2954
2955         spin_lock(&oi->ip_lock);
2956         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
2957         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
2958         spin_unlock(&oi->ip_lock);
2959
2960         ocfs2_dinode_new_extent_list(dir, di);
2961
2962         i_size_write(dir, sb->s_blocksize);
2963         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
2964
2965         di->i_size = cpu_to_le64(sb->s_blocksize);
2966         di->i_ctime = di->i_mtime = cpu_to_le64(inode_get_ctime_sec(dir));
2967         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode_get_ctime_nsec(dir));
2968         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2969
2970         /*
2971          * This should never fail as our extent list is empty and all
2972          * related blocks have been journaled already.
2973          */
2974         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
2975                                   0, NULL);
2976         if (ret) {
2977                 mlog_errno(ret);
2978                 goto out_commit;
2979         }
2980
2981         /*
2982          * Set i_blocks after the extent insert for the most up to
2983          * date ip_clusters value.
2984          */
2985         dir->i_blocks = ocfs2_inode_sector_count(dir);
2986
2987         ocfs2_journal_dirty(handle, di_bh);
2988
2989         if (ocfs2_supports_indexed_dirs(osb)) {
2990                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
2991                                                 dirdata_bh, meta_ac, dx_inline,
2992                                                 num_dx_entries, &dx_root_bh);
2993                 if (ret) {
2994                         mlog_errno(ret);
2995                         goto out_commit;
2996                 }
2997
2998                 if (dx_inline) {
2999                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3000                                                       dirdata_bh);
3001                 } else {
3002                         ocfs2_init_dx_root_extent_tree(&dx_et,
3003                                                        INODE_CACHE(dir),
3004                                                        dx_root_bh);
3005                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3006                                                   dx_insert_blkno, 1, 0, NULL);
3007                         if (ret)
3008                                 mlog_errno(ret);
3009                 }
3010         }
3011
3012         /*
3013          * We asked for two clusters, but only got one in the 1st
3014          * pass. Claim the 2nd cluster as a separate extent.
3015          */
3016         if (alloc > len) {
3017                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3018                                            &len);
3019                 if (ret) {
3020                         mlog_errno(ret);
3021                         goto out_commit;
3022                 }
3023                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3024
3025                 ret = ocfs2_insert_extent(handle, &et, 1,
3026                                           blkno, len, 0, NULL);
3027                 if (ret) {
3028                         mlog_errno(ret);
3029                         goto out_commit;
3030                 }
3031                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3032         }
3033
3034         *first_block_bh = dirdata_bh;
3035         dirdata_bh = NULL;
3036         if (ocfs2_supports_indexed_dirs(osb)) {
3037                 unsigned int off;
3038
3039                 if (!dx_inline) {
3040                         /*
3041                          * We need to return the correct block within the
3042                          * cluster which should hold our entry.
3043                          */
3044                         off = ocfs2_dx_dir_hash_idx(osb,
3045                                                     &lookup->dl_hinfo);
3046                         get_bh(dx_leaves[off]);
3047                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3048                 }
3049                 lookup->dl_dx_root_bh = dx_root_bh;
3050                 dx_root_bh = NULL;
3051         }
3052
3053 out_commit:
3054         if (ret < 0 && did_quota)
3055                 dquot_free_space_nodirty(dir, bytes_allocated);
3056
3057         ocfs2_commit_trans(osb, handle);
3058
3059 out:
3060         up_write(&oi->ip_alloc_sem);
3061         if (data_ac)
3062                 ocfs2_free_alloc_context(data_ac);
3063         if (meta_ac)
3064                 ocfs2_free_alloc_context(meta_ac);
3065
3066         if (dx_leaves) {
3067                 for (i = 0; i < num_dx_leaves; i++)
3068                         brelse(dx_leaves[i]);
3069                 kfree(dx_leaves);
3070         }
3071
3072         brelse(dirdata_bh);
3073         brelse(dx_root_bh);
3074
3075         return ret;
3076 }
3077
3078 /* returns a bh of the 1st new block in the allocation. */
3079 static int ocfs2_do_extend_dir(struct super_block *sb,
3080                                handle_t *handle,
3081                                struct inode *dir,
3082                                struct buffer_head *parent_fe_bh,
3083                                struct ocfs2_alloc_context *data_ac,
3084                                struct ocfs2_alloc_context *meta_ac,
3085                                struct buffer_head **new_bh)
3086 {
3087         int status;
3088         int extend, did_quota = 0;
3089         u64 p_blkno, v_blkno;
3090
3091         spin_lock(&OCFS2_I(dir)->ip_lock);
3092         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3093         spin_unlock(&OCFS2_I(dir)->ip_lock);
3094
3095         if (extend) {
3096                 u32 offset = OCFS2_I(dir)->ip_clusters;
3097
3098                 status = dquot_alloc_space_nodirty(dir,
3099                                         ocfs2_clusters_to_bytes(sb, 1));
3100                 if (status)
3101                         goto bail;
3102                 did_quota = 1;
3103
3104                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3105                                               1, 0, parent_fe_bh, handle,
3106                                               data_ac, meta_ac, NULL);
3107                 BUG_ON(status == -EAGAIN);
3108                 if (status < 0) {
3109                         mlog_errno(status);
3110                         goto bail;
3111                 }
3112         }
3113
3114         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3115         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3116         if (status < 0) {
3117                 mlog_errno(status);
3118                 goto bail;
3119         }
3120
3121         *new_bh = sb_getblk(sb, p_blkno);
3122         if (!*new_bh) {
3123                 status = -ENOMEM;
3124                 mlog_errno(status);
3125                 goto bail;
3126         }
3127         status = 0;
3128 bail:
3129         if (did_quota && status < 0)
3130                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3131         return status;
3132 }
3133
3134 /*
3135  * Assumes you already have a cluster lock on the directory.
3136  *
3137  * 'blocks_wanted' is only used if we have an inline directory which
3138  * is to be turned into an extent based one. The size of the dirent to
3139  * insert might be larger than the space gained by growing to just one
3140  * block, so we may have to grow the inode by two blocks in that case.
3141  *
3142  * If the directory is already indexed, dx_root_bh must be provided.
3143  */
3144 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3145                             struct inode *dir,
3146                             struct buffer_head *parent_fe_bh,
3147                             unsigned int blocks_wanted,
3148                             struct ocfs2_dir_lookup_result *lookup,
3149                             struct buffer_head **new_de_bh)
3150 {
3151         int status = 0;
3152         int credits, num_free_extents, drop_alloc_sem = 0;
3153         loff_t dir_i_size;
3154         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3155         struct ocfs2_extent_list *el = &fe->id2.i_list;
3156         struct ocfs2_alloc_context *data_ac = NULL;
3157         struct ocfs2_alloc_context *meta_ac = NULL;
3158         handle_t *handle = NULL;
3159         struct buffer_head *new_bh = NULL;
3160         struct ocfs2_dir_entry * de;
3161         struct super_block *sb = osb->sb;
3162         struct ocfs2_extent_tree et;
3163         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3164
3165         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3166                 /*
3167                  * This would be a code error as an inline directory should
3168                  * never have an index root.
3169                  */
3170                 BUG_ON(dx_root_bh);
3171
3172                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3173                                                  blocks_wanted, lookup,
3174                                                  &new_bh);
3175                 if (status) {
3176                         mlog_errno(status);
3177                         goto bail;
3178                 }
3179
3180                 /* Expansion from inline to an indexed directory will
3181                  * have given us this. */
3182                 dx_root_bh = lookup->dl_dx_root_bh;
3183
3184                 if (blocks_wanted == 1) {
3185                         /*
3186                          * If the new dirent will fit inside the space
3187                          * created by pushing out to one block, then
3188                          * we can complete the operation
3189                          * here. Otherwise we have to expand i_size
3190                          * and format the 2nd block below.
3191                          */
3192                         BUG_ON(new_bh == NULL);
3193                         goto bail_bh;
3194                 }
3195
3196                 /*
3197                  * Get rid of 'new_bh' - we want to format the 2nd
3198                  * data block and return that instead.
3199                  */
3200                 brelse(new_bh);
3201                 new_bh = NULL;
3202
3203                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3204                 drop_alloc_sem = 1;
3205                 dir_i_size = i_size_read(dir);
3206                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3207                 goto do_extend;
3208         }
3209
3210         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3211         drop_alloc_sem = 1;
3212         dir_i_size = i_size_read(dir);
3213         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3214                                dir_i_size);
3215
3216         /* dir->i_size is always block aligned. */
3217         spin_lock(&OCFS2_I(dir)->ip_lock);
3218         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3219                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3220                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3221                                               parent_fe_bh);
3222                 num_free_extents = ocfs2_num_free_extents(&et);
3223                 if (num_free_extents < 0) {
3224                         status = num_free_extents;
3225                         mlog_errno(status);
3226                         goto bail;
3227                 }
3228
3229                 if (!num_free_extents) {
3230                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3231                         if (status < 0) {
3232                                 if (status != -ENOSPC)
3233                                         mlog_errno(status);
3234                                 goto bail;
3235                         }
3236                 }
3237
3238                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3239                 if (status < 0) {
3240                         if (status != -ENOSPC)
3241                                 mlog_errno(status);
3242                         goto bail;
3243                 }
3244
3245                 if (ocfs2_dir_resv_allowed(osb))
3246                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3247
3248                 credits = ocfs2_calc_extend_credits(sb, el);
3249         } else {
3250                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3251                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3252         }
3253
3254 do_extend:
3255         if (ocfs2_dir_indexed(dir))
3256                 credits++; /* For attaching the new dirent block to the
3257                             * dx_root */
3258
3259         handle = ocfs2_start_trans(osb, credits);
3260         if (IS_ERR(handle)) {
3261                 status = PTR_ERR(handle);
3262                 handle = NULL;
3263                 mlog_errno(status);
3264                 goto bail;
3265         }
3266
3267         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3268                                      data_ac, meta_ac, &new_bh);
3269         if (status < 0) {
3270                 mlog_errno(status);
3271                 goto bail;
3272         }
3273
3274         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3275
3276         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3277                                          OCFS2_JOURNAL_ACCESS_CREATE);
3278         if (status < 0) {
3279                 mlog_errno(status);
3280                 goto bail;
3281         }
3282         memset(new_bh->b_data, 0, sb->s_blocksize);
3283
3284         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3285         de->inode = 0;
3286         if (ocfs2_supports_dir_trailer(dir)) {
3287                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3288
3289                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3290
3291                 if (ocfs2_dir_indexed(dir)) {
3292                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3293                                                            dx_root_bh, new_bh);
3294                         if (status) {
3295                                 mlog_errno(status);
3296                                 goto bail;
3297                         }
3298                 }
3299         } else {
3300                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3301         }
3302         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3303         ocfs2_journal_dirty(handle, new_bh);
3304
3305         dir_i_size += dir->i_sb->s_blocksize;
3306         i_size_write(dir, dir_i_size);
3307         dir->i_blocks = ocfs2_inode_sector_count(dir);
3308         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3309         if (status < 0) {
3310                 mlog_errno(status);
3311                 goto bail;
3312         }
3313
3314 bail_bh:
3315         *new_de_bh = new_bh;
3316         get_bh(*new_de_bh);
3317 bail:
3318         if (handle)
3319                 ocfs2_commit_trans(osb, handle);
3320         if (drop_alloc_sem)
3321                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3322
3323         if (data_ac)
3324                 ocfs2_free_alloc_context(data_ac);
3325         if (meta_ac)
3326                 ocfs2_free_alloc_context(meta_ac);
3327
3328         brelse(new_bh);
3329
3330         return status;
3331 }
3332
3333 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3334                                    const char *name, int namelen,
3335                                    struct buffer_head **ret_de_bh,
3336                                    unsigned int *blocks_wanted)
3337 {
3338         int ret;
3339         struct super_block *sb = dir->i_sb;
3340         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3341         struct ocfs2_dir_entry *de, *last_de = NULL;
3342         char *de_buf, *limit;
3343         unsigned long offset = 0;
3344         unsigned int rec_len, new_rec_len, free_space;
3345
3346         /*
3347          * This calculates how many free bytes we'd have in block zero, should
3348          * this function force expansion to an extent tree.
3349          */
3350         if (ocfs2_new_dir_wants_trailer(dir))
3351                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3352         else
3353                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3354
3355         de_buf = di->id2.i_data.id_data;
3356         limit = de_buf + i_size_read(dir);
3357         rec_len = OCFS2_DIR_REC_LEN(namelen);
3358
3359         while (de_buf < limit) {
3360                 de = (struct ocfs2_dir_entry *)de_buf;
3361
3362                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3363                         ret = -ENOENT;
3364                         goto out;
3365                 }
3366                 if (ocfs2_match(namelen, name, de)) {
3367                         ret = -EEXIST;
3368                         goto out;
3369                 }
3370                 /*
3371                  * No need to check for a trailing dirent record here as
3372                  * they're not used for inline dirs.
3373                  */
3374
3375                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3376                         /* Ok, we found a spot. Return this bh and let
3377                          * the caller actually fill it in. */
3378                         *ret_de_bh = di_bh;
3379                         get_bh(*ret_de_bh);
3380                         ret = 0;
3381                         goto out;
3382                 }
3383
3384                 last_de = de;
3385                 de_buf += le16_to_cpu(de->rec_len);
3386                 offset += le16_to_cpu(de->rec_len);
3387         }
3388
3389         /*
3390          * We're going to require expansion of the directory - figure
3391          * out how many blocks we'll need so that a place for the
3392          * dirent can be found.
3393          */
3394         *blocks_wanted = 1;
3395         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3396         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3397                 *blocks_wanted = 2;
3398
3399         ret = -ENOSPC;
3400 out:
3401         return ret;
3402 }
3403
3404 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3405                                    int namelen, struct buffer_head **ret_de_bh)
3406 {
3407         unsigned long offset;
3408         struct buffer_head *bh = NULL;
3409         unsigned short rec_len;
3410         struct ocfs2_dir_entry *de;
3411         struct super_block *sb = dir->i_sb;
3412         int status;
3413         int blocksize = dir->i_sb->s_blocksize;
3414
3415         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3416         if (status)
3417                 goto bail;
3418
3419         rec_len = OCFS2_DIR_REC_LEN(namelen);
3420         offset = 0;
3421         de = (struct ocfs2_dir_entry *) bh->b_data;
3422         while (1) {
3423                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3424                         brelse(bh);
3425                         bh = NULL;
3426
3427                         if (i_size_read(dir) <= offset) {
3428                                 /*
3429                                  * Caller will have to expand this
3430                                  * directory.
3431                                  */
3432                                 status = -ENOSPC;
3433                                 goto bail;
3434                         }
3435                         status = ocfs2_read_dir_block(dir,
3436                                              offset >> sb->s_blocksize_bits,
3437                                              &bh, 0);
3438                         if (status)
3439                                 goto bail;
3440
3441                         /* move to next block */
3442                         de = (struct ocfs2_dir_entry *) bh->b_data;
3443                 }
3444                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3445                         status = -ENOENT;
3446                         goto bail;
3447                 }
3448                 if (ocfs2_match(namelen, name, de)) {
3449                         status = -EEXIST;
3450                         goto bail;
3451                 }
3452
3453                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3454                                            blocksize))
3455                         goto next;
3456
3457                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3458                         /* Ok, we found a spot. Return this bh and let
3459                          * the caller actually fill it in. */
3460                         *ret_de_bh = bh;
3461                         get_bh(*ret_de_bh);
3462                         status = 0;
3463                         goto bail;
3464                 }
3465 next:
3466                 offset += le16_to_cpu(de->rec_len);
3467                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3468         }
3469
3470 bail:
3471         brelse(bh);
3472         if (status)
3473                 mlog_errno(status);
3474
3475         return status;
3476 }
3477
3478 static int dx_leaf_sort_cmp(const void *a, const void *b)
3479 {
3480         const struct ocfs2_dx_entry *entry1 = a;
3481         const struct ocfs2_dx_entry *entry2 = b;
3482         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3483         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3484         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3485         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3486
3487         if (major_hash1 > major_hash2)
3488                 return 1;
3489         if (major_hash1 < major_hash2)
3490                 return -1;
3491
3492         /*
3493          * It is not strictly necessary to sort by minor
3494          */
3495         if (minor_hash1 > minor_hash2)
3496                 return 1;
3497         if (minor_hash1 < minor_hash2)
3498                 return -1;
3499         return 0;
3500 }
3501
3502 static void dx_leaf_sort_swap(void *a, void *b, int size)
3503 {
3504         struct ocfs2_dx_entry *entry1 = a;
3505         struct ocfs2_dx_entry *entry2 = b;
3506
3507         BUG_ON(size != sizeof(*entry1));
3508
3509         swap(*entry1, *entry2);
3510 }
3511
3512 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3513 {
3514         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3515         int i, num = le16_to_cpu(dl_list->de_num_used);
3516
3517         for (i = 0; i < (num - 1); i++) {
3518                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3519                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3520                         return 0;
3521         }
3522
3523         return 1;
3524 }
3525
3526 /*
3527  * Find the optimal value to split this leaf on. This expects the leaf
3528  * entries to be in sorted order.
3529  *
3530  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3531  * the hash we want to insert.
3532  *
3533  * This function is only concerned with the major hash - that which
3534  * determines which cluster an item belongs to.
3535  */
3536 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3537                                         u32 leaf_cpos, u32 insert_hash,
3538                                         u32 *split_hash)
3539 {
3540         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3541         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3542         int allsame;
3543
3544         /*
3545          * There's a couple rare, but nasty corner cases we have to
3546          * check for here. All of them involve a leaf where all value
3547          * have the same hash, which is what we look for first.
3548          *
3549          * Most of the time, all of the above is false, and we simply
3550          * pick the median value for a split.
3551          */
3552         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3553         if (allsame) {
3554                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3555
3556                 if (val == insert_hash) {
3557                         /*
3558                          * No matter where we would choose to split,
3559                          * the new entry would want to occupy the same
3560                          * block as these. Since there's no space left
3561                          * in their existing block, we know there
3562                          * won't be space after the split.
3563                          */
3564                         return -ENOSPC;
3565                 }
3566
3567                 if (val == leaf_cpos) {
3568                         /*
3569                          * Because val is the same as leaf_cpos (which
3570                          * is the smallest value this leaf can have),
3571                          * yet is not equal to insert_hash, then we
3572                          * know that insert_hash *must* be larger than
3573                          * val (and leaf_cpos). At least cpos+1 in value.
3574                          *
3575                          * We also know then, that there cannot be an
3576                          * adjacent extent (otherwise we'd be looking
3577                          * at it). Choosing this value gives us a
3578                          * chance to get some contiguousness.
3579                          */
3580                         *split_hash = leaf_cpos + 1;
3581                         return 0;
3582                 }
3583
3584                 if (val > insert_hash) {
3585                         /*
3586                          * val can not be the same as insert hash, and
3587                          * also must be larger than leaf_cpos. Also,
3588                          * we know that there can't be a leaf between
3589                          * cpos and val, otherwise the entries with
3590                          * hash 'val' would be there.
3591                          */
3592                         *split_hash = val;
3593                         return 0;
3594                 }
3595
3596                 *split_hash = insert_hash;
3597                 return 0;
3598         }
3599
3600         /*
3601          * Since the records are sorted and the checks above
3602          * guaranteed that not all records in this block are the same,
3603          * we simple travel forward, from the median, and pick the 1st
3604          * record whose value is larger than leaf_cpos.
3605          */
3606         for (i = (num_used / 2); i < num_used; i++)
3607                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3608                     leaf_cpos)
3609                         break;
3610
3611         BUG_ON(i == num_used); /* Should be impossible */
3612         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3613         return 0;
3614 }
3615
3616 /*
3617  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3618  * larger than split_hash into new_dx_leaves. We use a temporary
3619  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3620  *
3621  * Since the block offset inside a leaf (cluster) is a constant mask
3622  * of minor_hash, we can optimize - an item at block offset X within
3623  * the original cluster, will be at offset X within the new cluster.
3624  */
3625 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3626                                        handle_t *handle,
3627                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3628                                        struct buffer_head **orig_dx_leaves,
3629                                        struct buffer_head **new_dx_leaves,
3630                                        int num_dx_leaves)
3631 {
3632         int i, j, num_used;
3633         u32 major_hash;
3634         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3635         struct ocfs2_dx_entry_list *orig_list, *tmp_list;
3636         struct ocfs2_dx_entry *dx_entry;
3637
3638         tmp_list = &tmp_dx_leaf->dl_list;
3639
3640         for (i = 0; i < num_dx_leaves; i++) {
3641                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3642                 orig_list = &orig_dx_leaf->dl_list;
3643                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3644
3645                 num_used = le16_to_cpu(orig_list->de_num_used);
3646
3647                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3648                 tmp_list->de_num_used = cpu_to_le16(0);
3649                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3650
3651                 for (j = 0; j < num_used; j++) {
3652                         dx_entry = &orig_list->de_entries[j];
3653                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3654                         if (major_hash >= split_hash)
3655                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3656                                                               dx_entry);
3657                         else
3658                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3659                                                               dx_entry);
3660                 }
3661                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3662
3663                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3664                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3665         }
3666 }
3667
3668 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3669                                           struct ocfs2_dx_root_block *dx_root)
3670 {
3671         int credits = ocfs2_clusters_to_blocks(osb->sb, 3);
3672
3673         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3674         credits += ocfs2_quota_trans_credits(osb->sb);
3675         return credits;
3676 }
3677
3678 /*
3679  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3680  * half our entries into.
3681  */
3682 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3683                                   struct buffer_head *dx_root_bh,
3684                                   struct buffer_head *dx_leaf_bh,
3685                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3686                                   u64 leaf_blkno)
3687 {
3688         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3689         int credits, ret, i, num_used, did_quota = 0;
3690         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3691         u64 orig_leaves_start;
3692         int num_dx_leaves;
3693         struct buffer_head **orig_dx_leaves = NULL;
3694         struct buffer_head **new_dx_leaves = NULL;
3695         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3696         struct ocfs2_extent_tree et;
3697         handle_t *handle = NULL;
3698         struct ocfs2_dx_root_block *dx_root;
3699         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3700
3701         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3702                                      (unsigned long long)leaf_blkno,
3703                                      insert_hash);
3704
3705         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3706
3707         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3708         /*
3709          * XXX: This is a rather large limit. We should use a more
3710          * realistic value.
3711          */
3712         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3713                 return -ENOSPC;
3714
3715         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3716         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3717                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3718                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3719                      (unsigned long long)leaf_blkno, num_used);
3720                 ret = -EIO;
3721                 goto out;
3722         }
3723
3724         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3725         if (!orig_dx_leaves) {
3726                 ret = -ENOMEM;
3727                 mlog_errno(ret);
3728                 goto out;
3729         }
3730
3731         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3732         if (!new_dx_leaves) {
3733                 ret = -ENOMEM;
3734                 mlog_errno(ret);
3735                 goto out;
3736         }
3737
3738         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3739         if (ret) {
3740                 if (ret != -ENOSPC)
3741                         mlog_errno(ret);
3742                 goto out;
3743         }
3744
3745         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3746         handle = ocfs2_start_trans(osb, credits);
3747         if (IS_ERR(handle)) {
3748                 ret = PTR_ERR(handle);
3749                 handle = NULL;
3750                 mlog_errno(ret);
3751                 goto out;
3752         }
3753
3754         ret = dquot_alloc_space_nodirty(dir,
3755                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3756         if (ret)
3757                 goto out_commit;
3758         did_quota = 1;
3759
3760         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3761                                       OCFS2_JOURNAL_ACCESS_WRITE);
3762         if (ret) {
3763                 mlog_errno(ret);
3764                 goto out_commit;
3765         }
3766
3767         /*
3768          * This block is changing anyway, so we can sort it in place.
3769          */
3770         sort(dx_leaf->dl_list.de_entries, num_used,
3771              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3772              dx_leaf_sort_swap);
3773
3774         ocfs2_journal_dirty(handle, dx_leaf_bh);
3775
3776         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3777                                            &split_hash);
3778         if (ret) {
3779                 mlog_errno(ret);
3780                 goto  out_commit;
3781         }
3782
3783         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3784
3785         /*
3786          * We have to carefully order operations here. There are items
3787          * which want to be in the new cluster before insert, but in
3788          * order to put those items in the new cluster, we alter the
3789          * old cluster. A failure to insert gets nasty.
3790          *
3791          * So, start by reserving writes to the old
3792          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3793          * the new cluster for us, before inserting it. The insert
3794          * won't happen if there's an error before that. Once the
3795          * insert is done then, we can transfer from one leaf into the
3796          * other without fear of hitting any error.
3797          */
3798
3799         /*
3800          * The leaf transfer wants some scratch space so that we don't
3801          * wind up doing a bunch of expensive memmove().
3802          */
3803         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3804         if (!tmp_dx_leaf) {
3805                 ret = -ENOMEM;
3806                 mlog_errno(ret);
3807                 goto out_commit;
3808         }
3809
3810         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3811         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3812                                    orig_dx_leaves);
3813         if (ret) {
3814                 mlog_errno(ret);
3815                 goto out_commit;
3816         }
3817
3818         cpos = split_hash;
3819         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3820                                        data_ac, meta_ac, new_dx_leaves,
3821                                        num_dx_leaves);
3822         if (ret) {
3823                 mlog_errno(ret);
3824                 goto out_commit;
3825         }
3826
3827         for (i = 0; i < num_dx_leaves; i++) {
3828                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3829                                               orig_dx_leaves[i],
3830                                               OCFS2_JOURNAL_ACCESS_WRITE);
3831                 if (ret) {
3832                         mlog_errno(ret);
3833                         goto out_commit;
3834                 }
3835
3836                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3837                                               new_dx_leaves[i],
3838                                               OCFS2_JOURNAL_ACCESS_WRITE);
3839                 if (ret) {
3840                         mlog_errno(ret);
3841                         goto out_commit;
3842                 }
3843         }
3844
3845         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3846                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3847
3848 out_commit:
3849         if (ret < 0 && did_quota)
3850                 dquot_free_space_nodirty(dir,
3851                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3852
3853         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3854         ocfs2_commit_trans(osb, handle);
3855
3856 out:
3857         if (orig_dx_leaves || new_dx_leaves) {
3858                 for (i = 0; i < num_dx_leaves; i++) {
3859                         if (orig_dx_leaves)
3860                                 brelse(orig_dx_leaves[i]);
3861                         if (new_dx_leaves)
3862                                 brelse(new_dx_leaves[i]);
3863                 }
3864                 kfree(orig_dx_leaves);
3865                 kfree(new_dx_leaves);
3866         }
3867
3868         if (meta_ac)
3869                 ocfs2_free_alloc_context(meta_ac);
3870         if (data_ac)
3871                 ocfs2_free_alloc_context(data_ac);
3872
3873         kfree(tmp_dx_leaf);
3874         return ret;
3875 }
3876
3877 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3878                                    struct buffer_head *di_bh,
3879                                    struct buffer_head *dx_root_bh,
3880                                    const char *name, int namelen,
3881                                    struct ocfs2_dir_lookup_result *lookup)
3882 {
3883         int ret, rebalanced = 0;
3884         struct ocfs2_dx_root_block *dx_root;
3885         struct buffer_head *dx_leaf_bh = NULL;
3886         struct ocfs2_dx_leaf *dx_leaf;
3887         u64 blkno;
3888         u32 leaf_cpos;
3889
3890         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3891
3892 restart_search:
3893         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3894                                   &leaf_cpos, &blkno);
3895         if (ret) {
3896                 mlog_errno(ret);
3897                 goto out;
3898         }
3899
3900         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3901         if (ret) {
3902                 mlog_errno(ret);
3903                 goto out;
3904         }
3905
3906         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3907
3908         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3909             le16_to_cpu(dx_leaf->dl_list.de_count)) {
3910                 if (rebalanced) {
3911                         /*
3912                          * Rebalancing should have provided us with
3913                          * space in an appropriate leaf.
3914                          *
3915                          * XXX: Is this an abnormal condition then?
3916                          * Should we print a message here?
3917                          */
3918                         ret = -ENOSPC;
3919                         goto out;
3920                 }
3921
3922                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3923                                              &lookup->dl_hinfo, leaf_cpos,
3924                                              blkno);
3925                 if (ret) {
3926                         if (ret != -ENOSPC)
3927                                 mlog_errno(ret);
3928                         goto out;
3929                 }
3930
3931                 /*
3932                  * Restart the lookup. The rebalance might have
3933                  * changed which block our item fits into. Mark our
3934                  * progress, so we only execute this once.
3935                  */
3936                 brelse(dx_leaf_bh);
3937                 dx_leaf_bh = NULL;
3938                 rebalanced = 1;
3939                 goto restart_search;
3940         }
3941
3942         lookup->dl_dx_leaf_bh = dx_leaf_bh;
3943         dx_leaf_bh = NULL;
3944
3945 out:
3946         brelse(dx_leaf_bh);
3947         return ret;
3948 }
3949
3950 static int ocfs2_search_dx_free_list(struct inode *dir,
3951                                      struct buffer_head *dx_root_bh,
3952                                      int namelen,
3953                                      struct ocfs2_dir_lookup_result *lookup)
3954 {
3955         int ret = -ENOSPC;
3956         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
3957         struct ocfs2_dir_block_trailer *db;
3958         u64 next_block;
3959         int rec_len = OCFS2_DIR_REC_LEN(namelen);
3960         struct ocfs2_dx_root_block *dx_root;
3961
3962         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3963         next_block = le64_to_cpu(dx_root->dr_free_blk);
3964
3965         while (next_block) {
3966                 brelse(prev_leaf_bh);
3967                 prev_leaf_bh = leaf_bh;
3968                 leaf_bh = NULL;
3969
3970                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
3971                 if (ret) {
3972                         mlog_errno(ret);
3973                         goto out;
3974                 }
3975
3976                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
3977                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
3978                         lookup->dl_leaf_bh = leaf_bh;
3979                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
3980                         leaf_bh = NULL;
3981                         prev_leaf_bh = NULL;
3982                         break;
3983                 }
3984
3985                 next_block = le64_to_cpu(db->db_free_next);
3986         }
3987
3988         if (!next_block)
3989                 ret = -ENOSPC;
3990
3991 out:
3992
3993         brelse(leaf_bh);
3994         brelse(prev_leaf_bh);
3995         return ret;
3996 }
3997
3998 static int ocfs2_expand_inline_dx_root(struct inode *dir,
3999                                        struct buffer_head *dx_root_bh)
4000 {
4001         int ret, num_dx_leaves, i, j, did_quota = 0;
4002         struct buffer_head **dx_leaves = NULL;
4003         struct ocfs2_extent_tree et;
4004         u64 insert_blkno;
4005         struct ocfs2_alloc_context *data_ac = NULL;
4006         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4007         handle_t *handle = NULL;
4008         struct ocfs2_dx_root_block *dx_root;
4009         struct ocfs2_dx_entry_list *entry_list;
4010         struct ocfs2_dx_entry *dx_entry;
4011         struct ocfs2_dx_leaf *target_leaf;
4012
4013         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4014         if (ret) {
4015                 mlog_errno(ret);
4016                 goto out;
4017         }
4018
4019         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4020         if (!dx_leaves) {
4021                 ret = -ENOMEM;
4022                 mlog_errno(ret);
4023                 goto out;
4024         }
4025
4026         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4027         if (IS_ERR(handle)) {
4028                 ret = PTR_ERR(handle);
4029                 mlog_errno(ret);
4030                 goto out;
4031         }
4032
4033         ret = dquot_alloc_space_nodirty(dir,
4034                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4035         if (ret)
4036                 goto out_commit;
4037         did_quota = 1;
4038
4039         /*
4040          * We do this up front, before the allocation, so that a
4041          * failure to add the dx_root_bh to the journal won't result
4042          * us losing clusters.
4043          */
4044         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4045                                       OCFS2_JOURNAL_ACCESS_WRITE);
4046         if (ret) {
4047                 mlog_errno(ret);
4048                 goto out_commit;
4049         }
4050
4051         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4052                                          num_dx_leaves, &insert_blkno);
4053         if (ret) {
4054                 mlog_errno(ret);
4055                 goto out_commit;
4056         }
4057
4058         /*
4059          * Transfer the entries from our dx_root into the appropriate
4060          * block
4061          */
4062         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4063         entry_list = &dx_root->dr_entries;
4064
4065         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4066                 dx_entry = &entry_list->de_entries[i];
4067
4068                 j = __ocfs2_dx_dir_hash_idx(osb,
4069                                             le32_to_cpu(dx_entry->dx_minor_hash));
4070                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4071
4072                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4073
4074                 /* Each leaf has been passed to the journal already
4075                  * via __ocfs2_dx_dir_new_cluster() */
4076         }
4077
4078         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4079         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4080                offsetof(struct ocfs2_dx_root_block, dr_list));
4081         dx_root->dr_list.l_count =
4082                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4083
4084         /* This should never fail considering we start with an empty
4085          * dx_root. */
4086         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4087         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4088         if (ret)
4089                 mlog_errno(ret);
4090         did_quota = 0;
4091
4092         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4093         ocfs2_journal_dirty(handle, dx_root_bh);
4094
4095 out_commit:
4096         if (ret < 0 && did_quota)
4097                 dquot_free_space_nodirty(dir,
4098                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4099
4100         ocfs2_commit_trans(osb, handle);
4101
4102 out:
4103         if (data_ac)
4104                 ocfs2_free_alloc_context(data_ac);
4105
4106         if (dx_leaves) {
4107                 for (i = 0; i < num_dx_leaves; i++)
4108                         brelse(dx_leaves[i]);
4109                 kfree(dx_leaves);
4110         }
4111         return ret;
4112 }
4113
4114 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4115 {
4116         struct ocfs2_dx_root_block *dx_root;
4117         struct ocfs2_dx_entry_list *entry_list;
4118
4119         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4120         entry_list = &dx_root->dr_entries;
4121
4122         if (le16_to_cpu(entry_list->de_num_used) >=
4123             le16_to_cpu(entry_list->de_count))
4124                 return -ENOSPC;
4125
4126         return 0;
4127 }
4128
4129 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4130                                            struct buffer_head *di_bh,
4131                                            const char *name,
4132                                            int namelen,
4133                                            struct ocfs2_dir_lookup_result *lookup)
4134 {
4135         int ret, free_dx_root = 1;
4136         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4137         struct buffer_head *dx_root_bh = NULL;
4138         struct buffer_head *leaf_bh = NULL;
4139         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4140         struct ocfs2_dx_root_block *dx_root;
4141
4142         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4143         if (ret) {
4144                 mlog_errno(ret);
4145                 goto out;
4146         }
4147
4148         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4149         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4150                 ret = -ENOSPC;
4151                 mlog_errno(ret);
4152                 goto out;
4153         }
4154
4155         if (ocfs2_dx_root_inline(dx_root)) {
4156                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4157
4158                 if (ret == 0)
4159                         goto search_el;
4160
4161                 /*
4162                  * We ran out of room in the root block. Expand it to
4163                  * an extent, then allow ocfs2_find_dir_space_dx to do
4164                  * the rest.
4165                  */
4166                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4167                 if (ret) {
4168                         mlog_errno(ret);
4169                         goto out;
4170                 }
4171         }
4172
4173         /*
4174          * Insert preparation for an indexed directory is split into two
4175          * steps. The call to find_dir_space_dx reserves room in the index for
4176          * an additional item. If we run out of space there, it's a real error
4177          * we can't continue on.
4178          */
4179         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4180                                       namelen, lookup);
4181         if (ret) {
4182                 mlog_errno(ret);
4183                 goto out;
4184         }
4185
4186 search_el:
4187         /*
4188          * Next, we need to find space in the unindexed tree. This call
4189          * searches using the free space linked list. If the unindexed tree
4190          * lacks sufficient space, we'll expand it below. The expansion code
4191          * is smart enough to add any new blocks to the free space list.
4192          */
4193         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4194         if (ret && ret != -ENOSPC) {
4195                 mlog_errno(ret);
4196                 goto out;
4197         }
4198
4199         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4200         lookup->dl_dx_root_bh = dx_root_bh;
4201         free_dx_root = 0;
4202
4203         if (ret == -ENOSPC) {
4204                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4205
4206                 if (ret) {
4207                         mlog_errno(ret);
4208                         goto out;
4209                 }
4210
4211                 /*
4212                  * We make the assumption here that new leaf blocks are added
4213                  * to the front of our free list.
4214                  */
4215                 lookup->dl_prev_leaf_bh = NULL;
4216                 lookup->dl_leaf_bh = leaf_bh;
4217         }
4218
4219 out:
4220         if (free_dx_root)
4221                 brelse(dx_root_bh);
4222         return ret;
4223 }
4224
4225 /*
4226  * Get a directory ready for insert. Any directory allocation required
4227  * happens here. Success returns zero, and enough context in the dir
4228  * lookup result that ocfs2_add_entry() will be able complete the task
4229  * with minimal performance impact.
4230  */
4231 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4232                                  struct inode *dir,
4233                                  struct buffer_head *parent_fe_bh,
4234                                  const char *name,
4235                                  int namelen,
4236                                  struct ocfs2_dir_lookup_result *lookup)
4237 {
4238         int ret;
4239         unsigned int blocks_wanted = 1;
4240         struct buffer_head *bh = NULL;
4241
4242         trace_ocfs2_prepare_dir_for_insert(
4243                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4244
4245         /*
4246          * Do this up front to reduce confusion.
4247          *
4248          * The directory might start inline, then be turned into an
4249          * indexed one, in which case we'd need to hash deep inside
4250          * ocfs2_find_dir_space_id(). Since
4251          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4252          * done, there seems no point in spreading out the calls. We
4253          * can optimize away the case where the file system doesn't
4254          * support indexing.
4255          */
4256         if (ocfs2_supports_indexed_dirs(osb))
4257                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4258
4259         if (ocfs2_dir_indexed(dir)) {
4260                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4261                                                       name, namelen, lookup);
4262                 if (ret)
4263                         mlog_errno(ret);
4264                 goto out;
4265         }
4266
4267         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4268                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4269                                               namelen, &bh, &blocks_wanted);
4270         } else
4271                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4272
4273         if (ret && ret != -ENOSPC) {
4274                 mlog_errno(ret);
4275                 goto out;
4276         }
4277
4278         if (ret == -ENOSPC) {
4279                 /*
4280                  * We have to expand the directory to add this name.
4281                  */
4282                 BUG_ON(bh);
4283
4284                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4285                                        lookup, &bh);
4286                 if (ret) {
4287                         if (ret != -ENOSPC)
4288                                 mlog_errno(ret);
4289                         goto out;
4290                 }
4291
4292                 BUG_ON(!bh);
4293         }
4294
4295         lookup->dl_leaf_bh = bh;
4296         bh = NULL;
4297 out:
4298         brelse(bh);
4299         return ret;
4300 }
4301
4302 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4303                                      struct buffer_head *di_bh,
4304                                      struct buffer_head *dx_root_bh)
4305 {
4306         int ret;
4307         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4308         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4309         struct ocfs2_dx_root_block *dx_root;
4310         struct inode *dx_alloc_inode = NULL;
4311         struct buffer_head *dx_alloc_bh = NULL;
4312         handle_t *handle;
4313         u64 blk;
4314         u16 bit;
4315         u64 bg_blkno;
4316
4317         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4318
4319         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4320                                         EXTENT_ALLOC_SYSTEM_INODE,
4321                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4322         if (!dx_alloc_inode) {
4323                 ret = -ENOMEM;
4324                 mlog_errno(ret);
4325                 goto out;
4326         }
4327         inode_lock(dx_alloc_inode);
4328
4329         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4330         if (ret) {
4331                 mlog_errno(ret);
4332                 goto out_mutex;
4333         }
4334
4335         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4336         if (IS_ERR(handle)) {
4337                 ret = PTR_ERR(handle);
4338                 mlog_errno(ret);
4339                 goto out_unlock;
4340         }
4341
4342         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4343                                       OCFS2_JOURNAL_ACCESS_WRITE);
4344         if (ret) {
4345                 mlog_errno(ret);
4346                 goto out_commit;
4347         }
4348
4349         spin_lock(&OCFS2_I(dir)->ip_lock);
4350         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4351         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4352         spin_unlock(&OCFS2_I(dir)->ip_lock);
4353         di->i_dx_root = cpu_to_le64(0ULL);
4354         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4355
4356         ocfs2_journal_dirty(handle, di_bh);
4357
4358         blk = le64_to_cpu(dx_root->dr_blkno);
4359         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4360         if (dx_root->dr_suballoc_loc)
4361                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4362         else
4363                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4364         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4365                                        bit, bg_blkno, 1);
4366         if (ret)
4367                 mlog_errno(ret);
4368
4369 out_commit:
4370         ocfs2_commit_trans(osb, handle);
4371
4372 out_unlock:
4373         ocfs2_inode_unlock(dx_alloc_inode, 1);
4374
4375 out_mutex:
4376         inode_unlock(dx_alloc_inode);
4377         brelse(dx_alloc_bh);
4378 out:
4379         iput(dx_alloc_inode);
4380         return ret;
4381 }
4382
4383 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4384 {
4385         int ret;
4386         unsigned int clen;
4387         u32 major_hash = UINT_MAX, p_cpos, cpos;
4388         u64 blkno;
4389         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4390         struct buffer_head *dx_root_bh = NULL;
4391         struct ocfs2_dx_root_block *dx_root;
4392         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4393         struct ocfs2_cached_dealloc_ctxt dealloc;
4394         struct ocfs2_extent_tree et;
4395
4396         ocfs2_init_dealloc_ctxt(&dealloc);
4397
4398         if (!ocfs2_dir_indexed(dir))
4399                 return 0;
4400
4401         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4402         if (ret) {
4403                 mlog_errno(ret);
4404                 goto out;
4405         }
4406         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4407
4408         if (ocfs2_dx_root_inline(dx_root))
4409                 goto remove_index;
4410
4411         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4412
4413         /* XXX: What if dr_clusters is too large? */
4414         while (le32_to_cpu(dx_root->dr_clusters)) {
4415                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4416                                               major_hash, &cpos, &blkno, &clen);
4417                 if (ret) {
4418                         mlog_errno(ret);
4419                         goto out;
4420                 }
4421
4422                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4423
4424                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4425                                                &dealloc, 0, false);
4426                 if (ret) {
4427                         mlog_errno(ret);
4428                         goto out;
4429                 }
4430
4431                 if (cpos == 0)
4432                         break;
4433
4434                 major_hash = cpos - 1;
4435         }
4436
4437 remove_index:
4438         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4439         if (ret) {
4440                 mlog_errno(ret);
4441                 goto out;
4442         }
4443
4444         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4445 out:
4446         ocfs2_schedule_truncate_log_flush(osb, 1);
4447         ocfs2_run_deallocs(osb, &dealloc);
4448
4449         brelse(dx_root_bh);
4450         return ret;
4451 }