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 (!namelen)
1597                 return -EINVAL;
1598
1599         if (ocfs2_dir_indexed(dir)) {
1600                 struct buffer_head *bh;
1601
1602                 /*
1603                  * An indexed dir may require that we update the free space
1604                  * list. Reserve a write to the previous node in the list so
1605                  * that we don't fail later.
1606                  *
1607                  * XXX: This can be either a dx_root_block, or an unindexed
1608                  * directory tree leaf block.
1609                  */
1610                 if (ocfs2_free_list_at_root(lookup)) {
1611                         bh = lookup->dl_dx_root_bh;
1612                         retval = ocfs2_journal_access_dr(handle,
1613                                                  INODE_CACHE(dir), bh,
1614                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1615                 } else {
1616                         bh = lookup->dl_prev_leaf_bh;
1617                         retval = ocfs2_journal_access_db(handle,
1618                                                  INODE_CACHE(dir), bh,
1619                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1620                 }
1621                 if (retval) {
1622                         mlog_errno(retval);
1623                         return retval;
1624                 }
1625         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1626                 data_start = di->id2.i_data.id_data;
1627                 size = i_size_read(dir);
1628
1629                 BUG_ON(insert_bh != parent_fe_bh);
1630         }
1631
1632         rec_len = OCFS2_DIR_REC_LEN(namelen);
1633         offset = 0;
1634         de = (struct ocfs2_dir_entry *) data_start;
1635         while (1) {
1636                 BUG_ON((char *)de >= (size + data_start));
1637
1638                 /* These checks should've already been passed by the
1639                  * prepare function, but I guess we can leave them
1640                  * here anyway. */
1641                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1642                         retval = -ENOENT;
1643                         goto bail;
1644                 }
1645                 if (ocfs2_match(namelen, name, de)) {
1646                         retval = -EEXIST;
1647                         goto bail;
1648                 }
1649
1650                 /* We're guaranteed that we should have space, so we
1651                  * can't possibly have hit the trailer...right? */
1652                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1653                                 "Hit dir trailer trying to insert %.*s "
1654                                 "(namelen %d) into directory %llu.  "
1655                                 "offset is %lu, trailer offset is %d\n",
1656                                 namelen, name, namelen,
1657                                 (unsigned long long)parent_fe_bh->b_blocknr,
1658                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1659
1660                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1661                         inode_set_mtime_to_ts(dir,
1662                                               inode_set_ctime_current(dir));
1663                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1664                         if (retval < 0) {
1665                                 mlog_errno(retval);
1666                                 goto bail;
1667                         }
1668
1669                         if (insert_bh == parent_fe_bh)
1670                                 retval = ocfs2_journal_access_di(handle,
1671                                                                  INODE_CACHE(dir),
1672                                                                  insert_bh,
1673                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1674                         else {
1675                                 retval = ocfs2_journal_access_db(handle,
1676                                                                  INODE_CACHE(dir),
1677                                                                  insert_bh,
1678                                               OCFS2_JOURNAL_ACCESS_WRITE);
1679
1680                                 if (!retval && ocfs2_dir_indexed(dir))
1681                                         retval = ocfs2_dx_dir_insert(dir,
1682                                                                 handle,
1683                                                                 lookup);
1684                         }
1685
1686                         if (retval) {
1687                                 mlog_errno(retval);
1688                                 goto bail;
1689                         }
1690
1691                         /* By now the buffer is marked for journaling */
1692                         offset += le16_to_cpu(de->rec_len);
1693                         if (le64_to_cpu(de->inode)) {
1694                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1695                                         OCFS2_DIR_REC_LEN(de->name_len));
1696                                 de1->rec_len =
1697                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1698                                         OCFS2_DIR_REC_LEN(de->name_len));
1699                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1700                                 de = de1;
1701                         }
1702                         de->file_type = FT_UNKNOWN;
1703                         if (blkno) {
1704                                 de->inode = cpu_to_le64(blkno);
1705                                 ocfs2_set_de_type(de, inode->i_mode);
1706                         } else
1707                                 de->inode = 0;
1708                         de->name_len = namelen;
1709                         memcpy(de->name, name, namelen);
1710
1711                         if (ocfs2_dir_indexed(dir))
1712                                 ocfs2_recalc_free_list(dir, handle, lookup);
1713
1714                         inode_inc_iversion(dir);
1715                         ocfs2_journal_dirty(handle, insert_bh);
1716                         retval = 0;
1717                         goto bail;
1718                 }
1719
1720                 offset += le16_to_cpu(de->rec_len);
1721                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1722         }
1723
1724         /* when you think about it, the assert above should prevent us
1725          * from ever getting here. */
1726         retval = -ENOSPC;
1727 bail:
1728         if (retval)
1729                 mlog_errno(retval);
1730
1731         return retval;
1732 }
1733
1734 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1735                                     u64 *f_version,
1736                                     struct dir_context *ctx)
1737 {
1738         int ret, i;
1739         unsigned long offset = ctx->pos;
1740         struct buffer_head *di_bh = NULL;
1741         struct ocfs2_dinode *di;
1742         struct ocfs2_inline_data *data;
1743         struct ocfs2_dir_entry *de;
1744
1745         ret = ocfs2_read_inode_block(inode, &di_bh);
1746         if (ret) {
1747                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1748                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1749                 goto out;
1750         }
1751
1752         di = (struct ocfs2_dinode *)di_bh->b_data;
1753         data = &di->id2.i_data;
1754
1755         while (ctx->pos < i_size_read(inode)) {
1756                 /* If the dir block has changed since the last call to
1757                  * readdir(2), then we might be pointing to an invalid
1758                  * dirent right now.  Scan from the start of the block
1759                  * to make sure. */
1760                 if (!inode_eq_iversion(inode, *f_version)) {
1761                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1762                                 de = (struct ocfs2_dir_entry *)
1763                                         (data->id_data + i);
1764                                 /* It's too expensive to do a full
1765                                  * dirent test each time round this
1766                                  * loop, but we do have to test at
1767                                  * least that it is non-zero.  A
1768                                  * failure will be detected in the
1769                                  * dirent test below. */
1770                                 if (le16_to_cpu(de->rec_len) <
1771                                     OCFS2_DIR_REC_LEN(1))
1772                                         break;
1773                                 i += le16_to_cpu(de->rec_len);
1774                         }
1775                         ctx->pos = offset = i;
1776                         *f_version = inode_query_iversion(inode);
1777                 }
1778
1779                 de = (struct ocfs2_dir_entry *) (data->id_data + ctx->pos);
1780                 if (!ocfs2_check_dir_entry(inode, de, di_bh, ctx->pos)) {
1781                         /* On error, skip the f_pos to the end. */
1782                         ctx->pos = i_size_read(inode);
1783                         break;
1784                 }
1785                 offset += le16_to_cpu(de->rec_len);
1786                 if (le64_to_cpu(de->inode)) {
1787                         if (!dir_emit(ctx, de->name, de->name_len,
1788                                       le64_to_cpu(de->inode),
1789                                       fs_ftype_to_dtype(de->file_type)))
1790                                 goto out;
1791                 }
1792                 ctx->pos += le16_to_cpu(de->rec_len);
1793         }
1794 out:
1795         brelse(di_bh);
1796         return 0;
1797 }
1798
1799 /*
1800  * NOTE: This function can be called against unindexed directories,
1801  * and indexed ones.
1802  */
1803 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1804                                     u64 *f_version,
1805                                     struct dir_context *ctx,
1806                                     bool persist)
1807 {
1808         unsigned long offset, blk, last_ra_blk = 0;
1809         int i;
1810         struct buffer_head * bh, * tmp;
1811         struct ocfs2_dir_entry * de;
1812         struct super_block * sb = inode->i_sb;
1813         unsigned int ra_sectors = 16;
1814         int stored = 0;
1815
1816         bh = NULL;
1817
1818         offset = ctx->pos & (sb->s_blocksize - 1);
1819
1820         while (ctx->pos < i_size_read(inode)) {
1821                 blk = ctx->pos >> sb->s_blocksize_bits;
1822                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1823                         /* Skip the corrupt dirblock and keep trying */
1824                         ctx->pos += sb->s_blocksize - offset;
1825                         continue;
1826                 }
1827
1828                 /* The idea here is to begin with 8k read-ahead and to stay
1829                  * 4k ahead of our current position.
1830                  *
1831                  * TODO: Use the pagecache for this. We just need to
1832                  * make sure it's cluster-safe... */
1833                 if (!last_ra_blk
1834                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1835                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1836                              i > 0; i--) {
1837                                 tmp = NULL;
1838                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1839                                                           OCFS2_BH_READAHEAD))
1840                                         brelse(tmp);
1841                         }
1842                         last_ra_blk = blk;
1843                         ra_sectors = 8;
1844                 }
1845
1846                 /* If the dir block has changed since the last call to
1847                  * readdir(2), then we might be pointing to an invalid
1848                  * dirent right now.  Scan from the start of the block
1849                  * to make sure. */
1850                 if (!inode_eq_iversion(inode, *f_version)) {
1851                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1852                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1853                                 /* It's too expensive to do a full
1854                                  * dirent test each time round this
1855                                  * loop, but we do have to test at
1856                                  * least that it is non-zero.  A
1857                                  * failure will be detected in the
1858                                  * dirent test below. */
1859                                 if (le16_to_cpu(de->rec_len) <
1860                                     OCFS2_DIR_REC_LEN(1))
1861                                         break;
1862                                 i += le16_to_cpu(de->rec_len);
1863                         }
1864                         offset = i;
1865                         ctx->pos = (ctx->pos & ~(sb->s_blocksize - 1))
1866                                 | offset;
1867                         *f_version = inode_query_iversion(inode);
1868                 }
1869
1870                 while (ctx->pos < i_size_read(inode)
1871                        && offset < sb->s_blocksize) {
1872                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1873                         if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1874                                 /* On error, skip the f_pos to the
1875                                    next block. */
1876                                 ctx->pos = (ctx->pos | (sb->s_blocksize - 1)) + 1;
1877                                 break;
1878                         }
1879                         if (le64_to_cpu(de->inode)) {
1880                                 if (!dir_emit(ctx, de->name,
1881                                                 de->name_len,
1882                                                 le64_to_cpu(de->inode),
1883                                         fs_ftype_to_dtype(de->file_type))) {
1884                                         brelse(bh);
1885                                         return 0;
1886                                 }
1887                                 stored++;
1888                         }
1889                         offset += le16_to_cpu(de->rec_len);
1890                         ctx->pos += le16_to_cpu(de->rec_len);
1891                 }
1892                 offset = 0;
1893                 brelse(bh);
1894                 bh = NULL;
1895                 if (!persist && stored)
1896                         break;
1897         }
1898         return 0;
1899 }
1900
1901 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1902                                  struct dir_context *ctx,
1903                                  bool persist)
1904 {
1905         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1906                 return ocfs2_dir_foreach_blk_id(inode, f_version, ctx);
1907         return ocfs2_dir_foreach_blk_el(inode, f_version, ctx, persist);
1908 }
1909
1910 /*
1911  * This is intended to be called from inside other kernel functions,
1912  * so we fake some arguments.
1913  */
1914 int ocfs2_dir_foreach(struct inode *inode, struct dir_context *ctx)
1915 {
1916         u64 version = inode_query_iversion(inode);
1917         ocfs2_dir_foreach_blk(inode, &version, ctx, true);
1918         return 0;
1919 }
1920
1921 /*
1922  * ocfs2_readdir()
1923  *
1924  */
1925 int ocfs2_readdir(struct file *file, struct dir_context *ctx)
1926 {
1927         int error = 0;
1928         struct inode *inode = file_inode(file);
1929         int lock_level = 0;
1930
1931         trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
1932
1933         error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level, 1);
1934         if (lock_level && error >= 0) {
1935                 /* We release EX lock which used to update atime
1936                  * and get PR lock again to reduce contention
1937                  * on commonly accessed directories. */
1938                 ocfs2_inode_unlock(inode, 1);
1939                 lock_level = 0;
1940                 error = ocfs2_inode_lock(inode, NULL, 0);
1941         }
1942         if (error < 0) {
1943                 if (error != -ENOENT)
1944                         mlog_errno(error);
1945                 /* we haven't got any yet, so propagate the error. */
1946                 goto bail_nolock;
1947         }
1948
1949         error = ocfs2_dir_foreach_blk(inode, &file->f_version, ctx, false);
1950
1951         ocfs2_inode_unlock(inode, lock_level);
1952         if (error)
1953                 mlog_errno(error);
1954
1955 bail_nolock:
1956
1957         return error;
1958 }
1959
1960 /*
1961  * NOTE: this should always be called with parent dir i_rwsem taken.
1962  */
1963 int ocfs2_find_files_on_disk(const char *name,
1964                              int namelen,
1965                              u64 *blkno,
1966                              struct inode *inode,
1967                              struct ocfs2_dir_lookup_result *lookup)
1968 {
1969         int status = -ENOENT;
1970
1971         trace_ocfs2_find_files_on_disk(namelen, name, blkno,
1972                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1973
1974         status = ocfs2_find_entry(name, namelen, inode, lookup);
1975         if (status)
1976                 goto leave;
1977
1978         *blkno = le64_to_cpu(lookup->dl_entry->inode);
1979
1980         status = 0;
1981 leave:
1982
1983         return status;
1984 }
1985
1986 /*
1987  * Convenience function for callers which just want the block number
1988  * mapped to a name and don't require the full dirent info, etc.
1989  */
1990 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
1991                                int namelen, u64 *blkno)
1992 {
1993         int ret;
1994         struct ocfs2_dir_lookup_result lookup = { NULL, };
1995
1996         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
1997         ocfs2_free_dir_lookup_result(&lookup);
1998
1999         return ret;
2000 }
2001
2002 /* Check for a name within a directory.
2003  *
2004  * Return 0 if the name does not exist
2005  * Return -EEXIST if the directory contains the name
2006  *
2007  * Callers should have i_rwsem + a cluster lock on dir
2008  */
2009 int ocfs2_check_dir_for_entry(struct inode *dir,
2010                               const char *name,
2011                               int namelen)
2012 {
2013         int ret = 0;
2014         struct ocfs2_dir_lookup_result lookup = { NULL, };
2015
2016         trace_ocfs2_check_dir_for_entry(
2017                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2018
2019         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0) {
2020                 ret = -EEXIST;
2021                 mlog_errno(ret);
2022         }
2023
2024         ocfs2_free_dir_lookup_result(&lookup);
2025
2026         return ret;
2027 }
2028
2029 struct ocfs2_empty_dir_priv {
2030         struct dir_context ctx;
2031         unsigned seen_dot;
2032         unsigned seen_dot_dot;
2033         unsigned seen_other;
2034         unsigned dx_dir;
2035 };
2036 static bool ocfs2_empty_dir_filldir(struct dir_context *ctx, const char *name,
2037                                    int name_len, loff_t pos, u64 ino,
2038                                    unsigned type)
2039 {
2040         struct ocfs2_empty_dir_priv *p =
2041                 container_of(ctx, struct ocfs2_empty_dir_priv, ctx);
2042
2043         /*
2044          * Check the positions of "." and ".." records to be sure
2045          * they're in the correct place.
2046          *
2047          * Indexed directories don't need to proceed past the first
2048          * two entries, so we end the scan after seeing '..'. Despite
2049          * that, we allow the scan to proceed In the event that we
2050          * have a corrupted indexed directory (no dot or dot dot
2051          * entries). This allows us to double check for existing
2052          * entries which might not have been found in the index.
2053          */
2054         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2055                 p->seen_dot = 1;
2056                 return true;
2057         }
2058
2059         if (name_len == 2 && !strncmp("..", name, 2) &&
2060             pos == OCFS2_DIR_REC_LEN(1)) {
2061                 p->seen_dot_dot = 1;
2062
2063                 if (p->dx_dir && p->seen_dot)
2064                         return false;
2065
2066                 return true;
2067         }
2068
2069         p->seen_other = 1;
2070         return false;
2071 }
2072
2073 static int ocfs2_empty_dir_dx(struct inode *inode,
2074                               struct ocfs2_empty_dir_priv *priv)
2075 {
2076         int ret;
2077         struct buffer_head *di_bh = NULL;
2078         struct buffer_head *dx_root_bh = NULL;
2079         struct ocfs2_dinode *di;
2080         struct ocfs2_dx_root_block *dx_root;
2081
2082         priv->dx_dir = 1;
2083
2084         ret = ocfs2_read_inode_block(inode, &di_bh);
2085         if (ret) {
2086                 mlog_errno(ret);
2087                 goto out;
2088         }
2089         di = (struct ocfs2_dinode *)di_bh->b_data;
2090
2091         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2092         if (ret) {
2093                 mlog_errno(ret);
2094                 goto out;
2095         }
2096         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2097
2098         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2099                 priv->seen_other = 1;
2100
2101 out:
2102         brelse(di_bh);
2103         brelse(dx_root_bh);
2104         return ret;
2105 }
2106
2107 /*
2108  * routine to check that the specified directory is empty (for rmdir)
2109  *
2110  * Returns 1 if dir is empty, zero otherwise.
2111  *
2112  * XXX: This is a performance problem for unindexed directories.
2113  */
2114 int ocfs2_empty_dir(struct inode *inode)
2115 {
2116         int ret;
2117         struct ocfs2_empty_dir_priv priv = {
2118                 .ctx.actor = ocfs2_empty_dir_filldir,
2119         };
2120
2121         if (ocfs2_dir_indexed(inode)) {
2122                 ret = ocfs2_empty_dir_dx(inode, &priv);
2123                 if (ret)
2124                         mlog_errno(ret);
2125                 /*
2126                  * We still run ocfs2_dir_foreach to get the checks
2127                  * for "." and "..".
2128                  */
2129         }
2130
2131         ret = ocfs2_dir_foreach(inode, &priv.ctx);
2132         if (ret)
2133                 mlog_errno(ret);
2134
2135         if (!priv.seen_dot || !priv.seen_dot_dot) {
2136                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2137                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2138                 /*
2139                  * XXX: Is it really safe to allow an unlink to continue?
2140                  */
2141                 return 1;
2142         }
2143
2144         return !priv.seen_other;
2145 }
2146
2147 /*
2148  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2149  * "..", which might be used during creation of a directory with a trailing
2150  * header. It is otherwise safe to ignore the return code.
2151  */
2152 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2153                                                           struct inode *parent,
2154                                                           char *start,
2155                                                           unsigned int size)
2156 {
2157         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2158
2159         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2160         de->name_len = 1;
2161         de->rec_len =
2162                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2163         strcpy(de->name, ".");
2164         ocfs2_set_de_type(de, S_IFDIR);
2165
2166         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2167         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2168         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2169         de->name_len = 2;
2170         strcpy(de->name, "..");
2171         ocfs2_set_de_type(de, S_IFDIR);
2172
2173         return de;
2174 }
2175
2176 /*
2177  * This works together with code in ocfs2_mknod_locked() which sets
2178  * the inline-data flag and initializes the inline-data section.
2179  */
2180 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2181                                  handle_t *handle,
2182                                  struct inode *parent,
2183                                  struct inode *inode,
2184                                  struct buffer_head *di_bh)
2185 {
2186         int ret;
2187         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2188         struct ocfs2_inline_data *data = &di->id2.i_data;
2189         unsigned int size = le16_to_cpu(data->id_count);
2190
2191         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2192                                       OCFS2_JOURNAL_ACCESS_WRITE);
2193         if (ret) {
2194                 mlog_errno(ret);
2195                 goto out;
2196         }
2197
2198         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2199         ocfs2_journal_dirty(handle, di_bh);
2200
2201         i_size_write(inode, size);
2202         set_nlink(inode, 2);
2203         inode->i_blocks = ocfs2_inode_sector_count(inode);
2204
2205         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2206         if (ret < 0)
2207                 mlog_errno(ret);
2208
2209 out:
2210         return ret;
2211 }
2212
2213 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2214                                  handle_t *handle,
2215                                  struct inode *parent,
2216                                  struct inode *inode,
2217                                  struct buffer_head *fe_bh,
2218                                  struct ocfs2_alloc_context *data_ac,
2219                                  struct buffer_head **ret_new_bh)
2220 {
2221         int status;
2222         unsigned int size = osb->sb->s_blocksize;
2223         struct buffer_head *new_bh = NULL;
2224         struct ocfs2_dir_entry *de;
2225
2226         if (ocfs2_new_dir_wants_trailer(inode))
2227                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2228
2229         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2230                                      data_ac, NULL, &new_bh);
2231         if (status < 0) {
2232                 mlog_errno(status);
2233                 goto bail;
2234         }
2235
2236         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2237
2238         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2239                                          OCFS2_JOURNAL_ACCESS_CREATE);
2240         if (status < 0) {
2241                 mlog_errno(status);
2242                 goto bail;
2243         }
2244         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2245
2246         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2247         if (ocfs2_new_dir_wants_trailer(inode)) {
2248                 int size = le16_to_cpu(de->rec_len);
2249
2250                 /*
2251                  * Figure out the size of the hole left over after
2252                  * insertion of '.' and '..'. The trailer wants this
2253                  * information.
2254                  */
2255                 size -= OCFS2_DIR_REC_LEN(2);
2256                 size -= sizeof(struct ocfs2_dir_block_trailer);
2257
2258                 ocfs2_init_dir_trailer(inode, new_bh, size);
2259         }
2260
2261         ocfs2_journal_dirty(handle, new_bh);
2262
2263         i_size_write(inode, inode->i_sb->s_blocksize);
2264         set_nlink(inode, 2);
2265         inode->i_blocks = ocfs2_inode_sector_count(inode);
2266         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2267         if (status < 0) {
2268                 mlog_errno(status);
2269                 goto bail;
2270         }
2271
2272         status = 0;
2273         if (ret_new_bh) {
2274                 *ret_new_bh = new_bh;
2275                 new_bh = NULL;
2276         }
2277 bail:
2278         brelse(new_bh);
2279
2280         return status;
2281 }
2282
2283 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2284                                      handle_t *handle, struct inode *dir,
2285                                      struct buffer_head *di_bh,
2286                                      struct buffer_head *dirdata_bh,
2287                                      struct ocfs2_alloc_context *meta_ac,
2288                                      int dx_inline, u32 num_entries,
2289                                      struct buffer_head **ret_dx_root_bh)
2290 {
2291         int ret;
2292         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2293         u16 dr_suballoc_bit;
2294         u64 suballoc_loc, dr_blkno;
2295         unsigned int num_bits;
2296         struct buffer_head *dx_root_bh = NULL;
2297         struct ocfs2_dx_root_block *dx_root;
2298         struct ocfs2_dir_block_trailer *trailer =
2299                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2300
2301         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2302                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2303         if (ret) {
2304                 mlog_errno(ret);
2305                 goto out;
2306         }
2307
2308         trace_ocfs2_dx_dir_attach_index(
2309                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2310                                 (unsigned long long)dr_blkno);
2311
2312         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2313         if (dx_root_bh == NULL) {
2314                 ret = -ENOMEM;
2315                 goto out;
2316         }
2317         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2318
2319         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2320                                       OCFS2_JOURNAL_ACCESS_CREATE);
2321         if (ret < 0) {
2322                 mlog_errno(ret);
2323                 goto out;
2324         }
2325
2326         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2327         memset(dx_root, 0, osb->sb->s_blocksize);
2328         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2329         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2330         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2331         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2332         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2333         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2334         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2335         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2336         if (le16_to_cpu(trailer->db_free_rec_len))
2337                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2338         else
2339                 dx_root->dr_free_blk = cpu_to_le64(0);
2340
2341         if (dx_inline) {
2342                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2343                 dx_root->dr_entries.de_count =
2344                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2345         } else {
2346                 dx_root->dr_list.l_count =
2347                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2348         }
2349         ocfs2_journal_dirty(handle, dx_root_bh);
2350
2351         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2352                                       OCFS2_JOURNAL_ACCESS_CREATE);
2353         if (ret) {
2354                 mlog_errno(ret);
2355                 goto out;
2356         }
2357
2358         di->i_dx_root = cpu_to_le64(dr_blkno);
2359
2360         spin_lock(&OCFS2_I(dir)->ip_lock);
2361         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2362         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2363         spin_unlock(&OCFS2_I(dir)->ip_lock);
2364
2365         ocfs2_journal_dirty(handle, di_bh);
2366
2367         *ret_dx_root_bh = dx_root_bh;
2368         dx_root_bh = NULL;
2369
2370 out:
2371         brelse(dx_root_bh);
2372         return ret;
2373 }
2374
2375 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2376                                        handle_t *handle, struct inode *dir,
2377                                        struct buffer_head **dx_leaves,
2378                                        int num_dx_leaves, u64 start_blk)
2379 {
2380         int ret, i;
2381         struct ocfs2_dx_leaf *dx_leaf;
2382         struct buffer_head *bh;
2383
2384         for (i = 0; i < num_dx_leaves; i++) {
2385                 bh = sb_getblk(osb->sb, start_blk + i);
2386                 if (bh == NULL) {
2387                         ret = -ENOMEM;
2388                         goto out;
2389                 }
2390                 dx_leaves[i] = bh;
2391
2392                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2393
2394                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2395                                               OCFS2_JOURNAL_ACCESS_CREATE);
2396                 if (ret < 0) {
2397                         mlog_errno(ret);
2398                         goto out;
2399                 }
2400
2401                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2402
2403                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2404                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2405                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2406                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2407                 dx_leaf->dl_list.de_count =
2408                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2409
2410                 trace_ocfs2_dx_dir_format_cluster(
2411                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2412                                 (unsigned long long)bh->b_blocknr,
2413                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2414
2415                 ocfs2_journal_dirty(handle, bh);
2416         }
2417
2418         ret = 0;
2419 out:
2420         return ret;
2421 }
2422
2423 /*
2424  * Allocates and formats a new cluster for use in an indexed dir
2425  * leaf. This version will not do the extent insert, so that it can be
2426  * used by operations which need careful ordering.
2427  */
2428 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2429                                       u32 cpos, handle_t *handle,
2430                                       struct ocfs2_alloc_context *data_ac,
2431                                       struct buffer_head **dx_leaves,
2432                                       int num_dx_leaves, u64 *ret_phys_blkno)
2433 {
2434         int ret;
2435         u32 phys, num;
2436         u64 phys_blkno;
2437         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2438
2439         /*
2440          * XXX: For create, this should claim cluster for the index
2441          * *before* the unindexed insert so that we have a better
2442          * chance of contiguousness as the directory grows in number
2443          * of entries.
2444          */
2445         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2446         if (ret) {
2447                 mlog_errno(ret);
2448                 goto out;
2449         }
2450
2451         /*
2452          * Format the new cluster first. That way, we're inserting
2453          * valid data.
2454          */
2455         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2456         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2457                                           num_dx_leaves, phys_blkno);
2458         if (ret) {
2459                 mlog_errno(ret);
2460                 goto out;
2461         }
2462
2463         *ret_phys_blkno = phys_blkno;
2464 out:
2465         return ret;
2466 }
2467
2468 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2469                                     struct ocfs2_extent_tree *et,
2470                                     u32 cpos, handle_t *handle,
2471                                     struct ocfs2_alloc_context *data_ac,
2472                                     struct ocfs2_alloc_context *meta_ac,
2473                                     struct buffer_head **dx_leaves,
2474                                     int num_dx_leaves)
2475 {
2476         int ret;
2477         u64 phys_blkno;
2478
2479         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2480                                          num_dx_leaves, &phys_blkno);
2481         if (ret) {
2482                 mlog_errno(ret);
2483                 goto out;
2484         }
2485
2486         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2487                                   meta_ac);
2488         if (ret)
2489                 mlog_errno(ret);
2490 out:
2491         return ret;
2492 }
2493
2494 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2495                                                         int *ret_num_leaves)
2496 {
2497         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2498         struct buffer_head **dx_leaves;
2499
2500         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2501                             GFP_NOFS);
2502         if (dx_leaves && ret_num_leaves)
2503                 *ret_num_leaves = num_dx_leaves;
2504
2505         return dx_leaves;
2506 }
2507
2508 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2509                                  handle_t *handle,
2510                                  struct inode *parent,
2511                                  struct inode *inode,
2512                                  struct buffer_head *di_bh,
2513                                  struct ocfs2_alloc_context *data_ac,
2514                                  struct ocfs2_alloc_context *meta_ac)
2515 {
2516         int ret;
2517         struct buffer_head *leaf_bh = NULL;
2518         struct buffer_head *dx_root_bh = NULL;
2519         struct ocfs2_dx_hinfo hinfo;
2520         struct ocfs2_dx_root_block *dx_root;
2521         struct ocfs2_dx_entry_list *entry_list;
2522
2523         /*
2524          * Our strategy is to create the directory as though it were
2525          * unindexed, then add the index block. This works with very
2526          * little complication since the state of a new directory is a
2527          * very well known quantity.
2528          *
2529          * Essentially, we have two dirents ("." and ".."), in the 1st
2530          * block which need indexing. These are easily inserted into
2531          * the index block.
2532          */
2533
2534         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2535                                     data_ac, &leaf_bh);
2536         if (ret) {
2537                 mlog_errno(ret);
2538                 goto out;
2539         }
2540
2541         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2542                                         meta_ac, 1, 2, &dx_root_bh);
2543         if (ret) {
2544                 mlog_errno(ret);
2545                 goto out;
2546         }
2547         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2548         entry_list = &dx_root->dr_entries;
2549
2550         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2551         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2552         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2553
2554         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2555         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2556
2557 out:
2558         brelse(dx_root_bh);
2559         brelse(leaf_bh);
2560         return ret;
2561 }
2562
2563 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2564                        handle_t *handle,
2565                        struct inode *parent,
2566                        struct inode *inode,
2567                        struct buffer_head *fe_bh,
2568                        struct ocfs2_alloc_context *data_ac,
2569                        struct ocfs2_alloc_context *meta_ac)
2570
2571 {
2572         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2573
2574         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2575                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2576
2577         if (ocfs2_supports_indexed_dirs(osb))
2578                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2579                                              data_ac, meta_ac);
2580
2581         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2582                                      data_ac, NULL);
2583 }
2584
2585 static int ocfs2_dx_dir_index_block(struct inode *dir,
2586                                     handle_t *handle,
2587                                     struct buffer_head **dx_leaves,
2588                                     int num_dx_leaves,
2589                                     u32 *num_dx_entries,
2590                                     struct buffer_head *dirent_bh)
2591 {
2592         int ret = 0, namelen, i;
2593         char *de_buf, *limit;
2594         struct ocfs2_dir_entry *de;
2595         struct buffer_head *dx_leaf_bh;
2596         struct ocfs2_dx_hinfo hinfo;
2597         u64 dirent_blk = dirent_bh->b_blocknr;
2598
2599         de_buf = dirent_bh->b_data;
2600         limit = de_buf + dir->i_sb->s_blocksize;
2601
2602         while (de_buf < limit) {
2603                 de = (struct ocfs2_dir_entry *)de_buf;
2604
2605                 namelen = de->name_len;
2606                 if (!namelen || !de->inode)
2607                         goto inc;
2608
2609                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2610
2611                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2612                 dx_leaf_bh = dx_leaves[i];
2613
2614                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2615                                                  dirent_blk, dx_leaf_bh);
2616                 if (ret) {
2617                         mlog_errno(ret);
2618                         goto out;
2619                 }
2620
2621                 *num_dx_entries = *num_dx_entries + 1;
2622
2623 inc:
2624                 de_buf += le16_to_cpu(de->rec_len);
2625         }
2626
2627 out:
2628         return ret;
2629 }
2630
2631 /*
2632  * XXX: This expects dx_root_bh to already be part of the transaction.
2633  */
2634 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2635                                          struct buffer_head *dx_root_bh,
2636                                          struct buffer_head *dirent_bh)
2637 {
2638         char *de_buf, *limit;
2639         struct ocfs2_dx_root_block *dx_root;
2640         struct ocfs2_dir_entry *de;
2641         struct ocfs2_dx_hinfo hinfo;
2642         u64 dirent_blk = dirent_bh->b_blocknr;
2643
2644         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2645
2646         de_buf = dirent_bh->b_data;
2647         limit = de_buf + dir->i_sb->s_blocksize;
2648
2649         while (de_buf < limit) {
2650                 de = (struct ocfs2_dir_entry *)de_buf;
2651
2652                 if (!de->name_len || !de->inode)
2653                         goto inc;
2654
2655                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2656
2657                 trace_ocfs2_dx_dir_index_root_block(
2658                                 (unsigned long long)dir->i_ino,
2659                                 hinfo.major_hash, hinfo.minor_hash,
2660                                 de->name_len, de->name,
2661                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2662
2663                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2664                                            dirent_blk);
2665
2666                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2667 inc:
2668                 de_buf += le16_to_cpu(de->rec_len);
2669         }
2670 }
2671
2672 /*
2673  * Count the number of inline directory entries in di_bh and compare
2674  * them against the number of entries we can hold in an inline dx root
2675  * block.
2676  */
2677 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2678                                          struct buffer_head *di_bh)
2679 {
2680         int dirent_count = 0;
2681         char *de_buf, *limit;
2682         struct ocfs2_dir_entry *de;
2683         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2684
2685         de_buf = di->id2.i_data.id_data;
2686         limit = de_buf + i_size_read(dir);
2687
2688         while (de_buf < limit) {
2689                 de = (struct ocfs2_dir_entry *)de_buf;
2690
2691                 if (de->name_len && de->inode)
2692                         dirent_count++;
2693
2694                 de_buf += le16_to_cpu(de->rec_len);
2695         }
2696
2697         /* We are careful to leave room for one extra record. */
2698         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2699 }
2700
2701 /*
2702  * Expand rec_len of the rightmost dirent in a directory block so that it
2703  * contains the end of our valid space for dirents. We do this during
2704  * expansion from an inline directory to one with extents. The first dir block
2705  * in that case is taken from the inline data portion of the inode block.
2706  *
2707  * This will also return the largest amount of contiguous space for a dirent
2708  * in the block. That value is *not* necessarily the last dirent, even after
2709  * expansion. The directory indexing code wants this value for free space
2710  * accounting. We do this here since we're already walking the entire dir
2711  * block.
2712  *
2713  * We add the dir trailer if this filesystem wants it.
2714  */
2715 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2716                                              struct inode *dir)
2717 {
2718         struct super_block *sb = dir->i_sb;
2719         struct ocfs2_dir_entry *de;
2720         struct ocfs2_dir_entry *prev_de;
2721         char *de_buf, *limit;
2722         unsigned int new_size = sb->s_blocksize;
2723         unsigned int bytes, this_hole;
2724         unsigned int largest_hole = 0;
2725
2726         if (ocfs2_new_dir_wants_trailer(dir))
2727                 new_size = ocfs2_dir_trailer_blk_off(sb);
2728
2729         bytes = new_size - old_size;
2730
2731         limit = start + old_size;
2732         de_buf = start;
2733         de = (struct ocfs2_dir_entry *)de_buf;
2734         do {
2735                 this_hole = ocfs2_figure_dirent_hole(de);
2736                 if (this_hole > largest_hole)
2737                         largest_hole = this_hole;
2738
2739                 prev_de = de;
2740                 de_buf += le16_to_cpu(de->rec_len);
2741                 de = (struct ocfs2_dir_entry *)de_buf;
2742         } while (de_buf < limit);
2743
2744         le16_add_cpu(&prev_de->rec_len, bytes);
2745
2746         /* We need to double check this after modification of the final
2747          * dirent. */
2748         this_hole = ocfs2_figure_dirent_hole(prev_de);
2749         if (this_hole > largest_hole)
2750                 largest_hole = this_hole;
2751
2752         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2753                 return largest_hole;
2754         return 0;
2755 }
2756
2757 /*
2758  * We allocate enough clusters to fulfill "blocks_wanted", but set
2759  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2760  * rest automatically for us.
2761  *
2762  * *first_block_bh is a pointer to the 1st data block allocated to the
2763  *  directory.
2764  */
2765 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2766                                    unsigned int blocks_wanted,
2767                                    struct ocfs2_dir_lookup_result *lookup,
2768                                    struct buffer_head **first_block_bh)
2769 {
2770         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2771         struct super_block *sb = dir->i_sb;
2772         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2773                 credits = ocfs2_inline_to_extents_credits(sb);
2774         u64 dx_insert_blkno, blkno,
2775                 bytes = blocks_wanted << sb->s_blocksize_bits;
2776         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2777         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2778         struct ocfs2_alloc_context *data_ac = NULL;
2779         struct ocfs2_alloc_context *meta_ac = NULL;
2780         struct buffer_head *dirdata_bh = NULL;
2781         struct buffer_head *dx_root_bh = NULL;
2782         struct buffer_head **dx_leaves = NULL;
2783         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2784         handle_t *handle;
2785         struct ocfs2_extent_tree et;
2786         struct ocfs2_extent_tree dx_et;
2787         int did_quota = 0, bytes_allocated = 0;
2788
2789         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2790
2791         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2792         dx_alloc = 0;
2793
2794         down_write(&oi->ip_alloc_sem);
2795
2796         if (ocfs2_supports_indexed_dirs(osb)) {
2797                 credits += ocfs2_add_dir_index_credits(sb);
2798
2799                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2800                 if (!dx_inline) {
2801                         /* Add one more cluster for an index leaf */
2802                         dx_alloc++;
2803                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2804                                                                 &num_dx_leaves);
2805                         if (!dx_leaves) {
2806                                 ret = -ENOMEM;
2807                                 mlog_errno(ret);
2808                                 goto out;
2809                         }
2810                 }
2811
2812                 /* This gets us the dx_root */
2813                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2814                 if (ret) {
2815                         mlog_errno(ret);
2816                         goto out;
2817                 }
2818         }
2819
2820         /*
2821          * We should never need more than 2 clusters for the unindexed
2822          * tree - maximum dirent size is far less than one block. In
2823          * fact, the only time we'd need more than one cluster is if
2824          * blocksize == clustersize and the dirent won't fit in the
2825          * extra space that the expansion to a single block gives. As
2826          * of today, that only happens on 4k/4k file systems.
2827          */
2828         BUG_ON(alloc > 2);
2829
2830         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2831         if (ret) {
2832                 mlog_errno(ret);
2833                 goto out;
2834         }
2835
2836         /*
2837          * Prepare for worst case allocation scenario of two separate
2838          * extents in the unindexed tree.
2839          */
2840         if (alloc == 2)
2841                 credits += OCFS2_SUBALLOC_ALLOC;
2842
2843         handle = ocfs2_start_trans(osb, credits);
2844         if (IS_ERR(handle)) {
2845                 ret = PTR_ERR(handle);
2846                 mlog_errno(ret);
2847                 goto out;
2848         }
2849
2850         ret = dquot_alloc_space_nodirty(dir,
2851                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2852         if (ret)
2853                 goto out_commit;
2854         did_quota = 1;
2855
2856         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2857                 /*
2858                  * Allocate our index cluster first, to maximize the
2859                  * possibility that unindexed leaves grow
2860                  * contiguously.
2861                  */
2862                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2863                                                  dx_leaves, num_dx_leaves,
2864                                                  &dx_insert_blkno);
2865                 if (ret) {
2866                         mlog_errno(ret);
2867                         goto out_commit;
2868                 }
2869                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2870         }
2871
2872         /*
2873          * Try to claim as many clusters as the bitmap can give though
2874          * if we only get one now, that's enough to continue. The rest
2875          * will be claimed after the conversion to extents.
2876          */
2877         if (ocfs2_dir_resv_allowed(osb))
2878                 data_ac->ac_resv = &oi->ip_la_data_resv;
2879         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2880         if (ret) {
2881                 mlog_errno(ret);
2882                 goto out_commit;
2883         }
2884         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2885
2886         /*
2887          * Operations are carefully ordered so that we set up the new
2888          * data block first. The conversion from inline data to
2889          * extents follows.
2890          */
2891         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2892         dirdata_bh = sb_getblk(sb, blkno);
2893         if (!dirdata_bh) {
2894                 ret = -ENOMEM;
2895                 mlog_errno(ret);
2896                 goto out_commit;
2897         }
2898
2899         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2900
2901         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2902                                       OCFS2_JOURNAL_ACCESS_CREATE);
2903         if (ret) {
2904                 mlog_errno(ret);
2905                 goto out_commit;
2906         }
2907
2908         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2909         memset(dirdata_bh->b_data + i_size_read(dir), 0,
2910                sb->s_blocksize - i_size_read(dir));
2911         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2912         if (ocfs2_new_dir_wants_trailer(dir)) {
2913                 /*
2914                  * Prepare the dir trailer up front. It will otherwise look
2915                  * like a valid dirent. Even if inserting the index fails
2916                  * (unlikely), then all we'll have done is given first dir
2917                  * block a small amount of fragmentation.
2918                  */
2919                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2920         }
2921
2922         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2923         ocfs2_journal_dirty(handle, dirdata_bh);
2924
2925         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2926                 /*
2927                  * Dx dirs with an external cluster need to do this up
2928                  * front. Inline dx root's get handled later, after
2929                  * we've allocated our root block. We get passed back
2930                  * a total number of items so that dr_num_entries can
2931                  * be correctly set once the dx_root has been
2932                  * allocated.
2933                  */
2934                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2935                                                num_dx_leaves, &num_dx_entries,
2936                                                dirdata_bh);
2937                 if (ret) {
2938                         mlog_errno(ret);
2939                         goto out_commit;
2940                 }
2941         }
2942
2943         /*
2944          * Set extent, i_size, etc on the directory. After this, the
2945          * inode should contain the same exact dirents as before and
2946          * be fully accessible from system calls.
2947          *
2948          * We let the later dirent insert modify c/mtime - to the user
2949          * the data hasn't changed.
2950          */
2951         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2952                                       OCFS2_JOURNAL_ACCESS_CREATE);
2953         if (ret) {
2954                 mlog_errno(ret);
2955                 goto out_commit;
2956         }
2957
2958         spin_lock(&oi->ip_lock);
2959         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
2960         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
2961         spin_unlock(&oi->ip_lock);
2962
2963         ocfs2_dinode_new_extent_list(dir, di);
2964
2965         i_size_write(dir, sb->s_blocksize);
2966         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
2967
2968         di->i_size = cpu_to_le64(sb->s_blocksize);
2969         di->i_ctime = di->i_mtime = cpu_to_le64(inode_get_ctime_sec(dir));
2970         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode_get_ctime_nsec(dir));
2971         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2972
2973         /*
2974          * This should never fail as our extent list is empty and all
2975          * related blocks have been journaled already.
2976          */
2977         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
2978                                   0, NULL);
2979         if (ret) {
2980                 mlog_errno(ret);
2981                 goto out_commit;
2982         }
2983
2984         /*
2985          * Set i_blocks after the extent insert for the most up to
2986          * date ip_clusters value.
2987          */
2988         dir->i_blocks = ocfs2_inode_sector_count(dir);
2989
2990         ocfs2_journal_dirty(handle, di_bh);
2991
2992         if (ocfs2_supports_indexed_dirs(osb)) {
2993                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
2994                                                 dirdata_bh, meta_ac, dx_inline,
2995                                                 num_dx_entries, &dx_root_bh);
2996                 if (ret) {
2997                         mlog_errno(ret);
2998                         goto out_commit;
2999                 }
3000
3001                 if (dx_inline) {
3002                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3003                                                       dirdata_bh);
3004                 } else {
3005                         ocfs2_init_dx_root_extent_tree(&dx_et,
3006                                                        INODE_CACHE(dir),
3007                                                        dx_root_bh);
3008                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3009                                                   dx_insert_blkno, 1, 0, NULL);
3010                         if (ret)
3011                                 mlog_errno(ret);
3012                 }
3013         }
3014
3015         /*
3016          * We asked for two clusters, but only got one in the 1st
3017          * pass. Claim the 2nd cluster as a separate extent.
3018          */
3019         if (alloc > len) {
3020                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3021                                            &len);
3022                 if (ret) {
3023                         mlog_errno(ret);
3024                         goto out_commit;
3025                 }
3026                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3027
3028                 ret = ocfs2_insert_extent(handle, &et, 1,
3029                                           blkno, len, 0, NULL);
3030                 if (ret) {
3031                         mlog_errno(ret);
3032                         goto out_commit;
3033                 }
3034                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3035         }
3036
3037         *first_block_bh = dirdata_bh;
3038         dirdata_bh = NULL;
3039         if (ocfs2_supports_indexed_dirs(osb)) {
3040                 unsigned int off;
3041
3042                 if (!dx_inline) {
3043                         /*
3044                          * We need to return the correct block within the
3045                          * cluster which should hold our entry.
3046                          */
3047                         off = ocfs2_dx_dir_hash_idx(osb,
3048                                                     &lookup->dl_hinfo);
3049                         get_bh(dx_leaves[off]);
3050                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3051                 }
3052                 lookup->dl_dx_root_bh = dx_root_bh;
3053                 dx_root_bh = NULL;
3054         }
3055
3056 out_commit:
3057         if (ret < 0 && did_quota)
3058                 dquot_free_space_nodirty(dir, bytes_allocated);
3059
3060         ocfs2_commit_trans(osb, handle);
3061
3062 out:
3063         up_write(&oi->ip_alloc_sem);
3064         if (data_ac)
3065                 ocfs2_free_alloc_context(data_ac);
3066         if (meta_ac)
3067                 ocfs2_free_alloc_context(meta_ac);
3068
3069         if (dx_leaves) {
3070                 for (i = 0; i < num_dx_leaves; i++)
3071                         brelse(dx_leaves[i]);
3072                 kfree(dx_leaves);
3073         }
3074
3075         brelse(dirdata_bh);
3076         brelse(dx_root_bh);
3077
3078         return ret;
3079 }
3080
3081 /* returns a bh of the 1st new block in the allocation. */
3082 static int ocfs2_do_extend_dir(struct super_block *sb,
3083                                handle_t *handle,
3084                                struct inode *dir,
3085                                struct buffer_head *parent_fe_bh,
3086                                struct ocfs2_alloc_context *data_ac,
3087                                struct ocfs2_alloc_context *meta_ac,
3088                                struct buffer_head **new_bh)
3089 {
3090         int status;
3091         int extend, did_quota = 0;
3092         u64 p_blkno, v_blkno;
3093
3094         spin_lock(&OCFS2_I(dir)->ip_lock);
3095         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3096         spin_unlock(&OCFS2_I(dir)->ip_lock);
3097
3098         if (extend) {
3099                 u32 offset = OCFS2_I(dir)->ip_clusters;
3100
3101                 status = dquot_alloc_space_nodirty(dir,
3102                                         ocfs2_clusters_to_bytes(sb, 1));
3103                 if (status)
3104                         goto bail;
3105                 did_quota = 1;
3106
3107                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3108                                               1, 0, parent_fe_bh, handle,
3109                                               data_ac, meta_ac, NULL);
3110                 BUG_ON(status == -EAGAIN);
3111                 if (status < 0) {
3112                         mlog_errno(status);
3113                         goto bail;
3114                 }
3115         }
3116
3117         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3118         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3119         if (status < 0) {
3120                 mlog_errno(status);
3121                 goto bail;
3122         }
3123
3124         *new_bh = sb_getblk(sb, p_blkno);
3125         if (!*new_bh) {
3126                 status = -ENOMEM;
3127                 mlog_errno(status);
3128                 goto bail;
3129         }
3130         status = 0;
3131 bail:
3132         if (did_quota && status < 0)
3133                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3134         return status;
3135 }
3136
3137 /*
3138  * Assumes you already have a cluster lock on the directory.
3139  *
3140  * 'blocks_wanted' is only used if we have an inline directory which
3141  * is to be turned into an extent based one. The size of the dirent to
3142  * insert might be larger than the space gained by growing to just one
3143  * block, so we may have to grow the inode by two blocks in that case.
3144  *
3145  * If the directory is already indexed, dx_root_bh must be provided.
3146  */
3147 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3148                             struct inode *dir,
3149                             struct buffer_head *parent_fe_bh,
3150                             unsigned int blocks_wanted,
3151                             struct ocfs2_dir_lookup_result *lookup,
3152                             struct buffer_head **new_de_bh)
3153 {
3154         int status = 0;
3155         int credits, num_free_extents, drop_alloc_sem = 0;
3156         loff_t dir_i_size;
3157         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3158         struct ocfs2_extent_list *el = &fe->id2.i_list;
3159         struct ocfs2_alloc_context *data_ac = NULL;
3160         struct ocfs2_alloc_context *meta_ac = NULL;
3161         handle_t *handle = NULL;
3162         struct buffer_head *new_bh = NULL;
3163         struct ocfs2_dir_entry * de;
3164         struct super_block *sb = osb->sb;
3165         struct ocfs2_extent_tree et;
3166         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3167
3168         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3169                 /*
3170                  * This would be a code error as an inline directory should
3171                  * never have an index root.
3172                  */
3173                 BUG_ON(dx_root_bh);
3174
3175                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3176                                                  blocks_wanted, lookup,
3177                                                  &new_bh);
3178                 if (status) {
3179                         mlog_errno(status);
3180                         goto bail;
3181                 }
3182
3183                 /* Expansion from inline to an indexed directory will
3184                  * have given us this. */
3185                 dx_root_bh = lookup->dl_dx_root_bh;
3186
3187                 if (blocks_wanted == 1) {
3188                         /*
3189                          * If the new dirent will fit inside the space
3190                          * created by pushing out to one block, then
3191                          * we can complete the operation
3192                          * here. Otherwise we have to expand i_size
3193                          * and format the 2nd block below.
3194                          */
3195                         BUG_ON(new_bh == NULL);
3196                         goto bail_bh;
3197                 }
3198
3199                 /*
3200                  * Get rid of 'new_bh' - we want to format the 2nd
3201                  * data block and return that instead.
3202                  */
3203                 brelse(new_bh);
3204                 new_bh = NULL;
3205
3206                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3207                 drop_alloc_sem = 1;
3208                 dir_i_size = i_size_read(dir);
3209                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3210                 goto do_extend;
3211         }
3212
3213         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3214         drop_alloc_sem = 1;
3215         dir_i_size = i_size_read(dir);
3216         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3217                                dir_i_size);
3218
3219         /* dir->i_size is always block aligned. */
3220         spin_lock(&OCFS2_I(dir)->ip_lock);
3221         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3222                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3223                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3224                                               parent_fe_bh);
3225                 num_free_extents = ocfs2_num_free_extents(&et);
3226                 if (num_free_extents < 0) {
3227                         status = num_free_extents;
3228                         mlog_errno(status);
3229                         goto bail;
3230                 }
3231
3232                 if (!num_free_extents) {
3233                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3234                         if (status < 0) {
3235                                 if (status != -ENOSPC)
3236                                         mlog_errno(status);
3237                                 goto bail;
3238                         }
3239                 }
3240
3241                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3242                 if (status < 0) {
3243                         if (status != -ENOSPC)
3244                                 mlog_errno(status);
3245                         goto bail;
3246                 }
3247
3248                 if (ocfs2_dir_resv_allowed(osb))
3249                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3250
3251                 credits = ocfs2_calc_extend_credits(sb, el);
3252         } else {
3253                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3254                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3255         }
3256
3257 do_extend:
3258         if (ocfs2_dir_indexed(dir))
3259                 credits++; /* For attaching the new dirent block to the
3260                             * dx_root */
3261
3262         handle = ocfs2_start_trans(osb, credits);
3263         if (IS_ERR(handle)) {
3264                 status = PTR_ERR(handle);
3265                 handle = NULL;
3266                 mlog_errno(status);
3267                 goto bail;
3268         }
3269
3270         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3271                                      data_ac, meta_ac, &new_bh);
3272         if (status < 0) {
3273                 mlog_errno(status);
3274                 goto bail;
3275         }
3276
3277         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3278
3279         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3280                                          OCFS2_JOURNAL_ACCESS_CREATE);
3281         if (status < 0) {
3282                 mlog_errno(status);
3283                 goto bail;
3284         }
3285         memset(new_bh->b_data, 0, sb->s_blocksize);
3286
3287         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3288         de->inode = 0;
3289         if (ocfs2_supports_dir_trailer(dir)) {
3290                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3291
3292                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3293
3294                 if (ocfs2_dir_indexed(dir)) {
3295                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3296                                                            dx_root_bh, new_bh);
3297                         if (status) {
3298                                 mlog_errno(status);
3299                                 goto bail;
3300                         }
3301                 }
3302         } else {
3303                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3304         }
3305         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3306         ocfs2_journal_dirty(handle, new_bh);
3307
3308         dir_i_size += dir->i_sb->s_blocksize;
3309         i_size_write(dir, dir_i_size);
3310         dir->i_blocks = ocfs2_inode_sector_count(dir);
3311         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3312         if (status < 0) {
3313                 mlog_errno(status);
3314                 goto bail;
3315         }
3316
3317 bail_bh:
3318         *new_de_bh = new_bh;
3319         get_bh(*new_de_bh);
3320 bail:
3321         if (handle)
3322                 ocfs2_commit_trans(osb, handle);
3323         if (drop_alloc_sem)
3324                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3325
3326         if (data_ac)
3327                 ocfs2_free_alloc_context(data_ac);
3328         if (meta_ac)
3329                 ocfs2_free_alloc_context(meta_ac);
3330
3331         brelse(new_bh);
3332
3333         return status;
3334 }
3335
3336 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3337                                    const char *name, int namelen,
3338                                    struct buffer_head **ret_de_bh,
3339                                    unsigned int *blocks_wanted)
3340 {
3341         int ret;
3342         struct super_block *sb = dir->i_sb;
3343         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3344         struct ocfs2_dir_entry *de, *last_de = NULL;
3345         char *de_buf, *limit;
3346         unsigned long offset = 0;
3347         unsigned int rec_len, new_rec_len, free_space;
3348
3349         /*
3350          * This calculates how many free bytes we'd have in block zero, should
3351          * this function force expansion to an extent tree.
3352          */
3353         if (ocfs2_new_dir_wants_trailer(dir))
3354                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3355         else
3356                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3357
3358         de_buf = di->id2.i_data.id_data;
3359         limit = de_buf + i_size_read(dir);
3360         rec_len = OCFS2_DIR_REC_LEN(namelen);
3361
3362         while (de_buf < limit) {
3363                 de = (struct ocfs2_dir_entry *)de_buf;
3364
3365                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3366                         ret = -ENOENT;
3367                         goto out;
3368                 }
3369                 if (ocfs2_match(namelen, name, de)) {
3370                         ret = -EEXIST;
3371                         goto out;
3372                 }
3373                 /*
3374                  * No need to check for a trailing dirent record here as
3375                  * they're not used for inline dirs.
3376                  */
3377
3378                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3379                         /* Ok, we found a spot. Return this bh and let
3380                          * the caller actually fill it in. */
3381                         *ret_de_bh = di_bh;
3382                         get_bh(*ret_de_bh);
3383                         ret = 0;
3384                         goto out;
3385                 }
3386
3387                 last_de = de;
3388                 de_buf += le16_to_cpu(de->rec_len);
3389                 offset += le16_to_cpu(de->rec_len);
3390         }
3391
3392         /*
3393          * We're going to require expansion of the directory - figure
3394          * out how many blocks we'll need so that a place for the
3395          * dirent can be found.
3396          */
3397         *blocks_wanted = 1;
3398         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3399         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3400                 *blocks_wanted = 2;
3401
3402         ret = -ENOSPC;
3403 out:
3404         return ret;
3405 }
3406
3407 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3408                                    int namelen, struct buffer_head **ret_de_bh)
3409 {
3410         unsigned long offset;
3411         struct buffer_head *bh = NULL;
3412         unsigned short rec_len;
3413         struct ocfs2_dir_entry *de;
3414         struct super_block *sb = dir->i_sb;
3415         int status;
3416         int blocksize = dir->i_sb->s_blocksize;
3417
3418         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3419         if (status)
3420                 goto bail;
3421
3422         rec_len = OCFS2_DIR_REC_LEN(namelen);
3423         offset = 0;
3424         de = (struct ocfs2_dir_entry *) bh->b_data;
3425         while (1) {
3426                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3427                         brelse(bh);
3428                         bh = NULL;
3429
3430                         if (i_size_read(dir) <= offset) {
3431                                 /*
3432                                  * Caller will have to expand this
3433                                  * directory.
3434                                  */
3435                                 status = -ENOSPC;
3436                                 goto bail;
3437                         }
3438                         status = ocfs2_read_dir_block(dir,
3439                                              offset >> sb->s_blocksize_bits,
3440                                              &bh, 0);
3441                         if (status)
3442                                 goto bail;
3443
3444                         /* move to next block */
3445                         de = (struct ocfs2_dir_entry *) bh->b_data;
3446                 }
3447                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3448                         status = -ENOENT;
3449                         goto bail;
3450                 }
3451                 if (ocfs2_match(namelen, name, de)) {
3452                         status = -EEXIST;
3453                         goto bail;
3454                 }
3455
3456                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3457                                            blocksize))
3458                         goto next;
3459
3460                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3461                         /* Ok, we found a spot. Return this bh and let
3462                          * the caller actually fill it in. */
3463                         *ret_de_bh = bh;
3464                         get_bh(*ret_de_bh);
3465                         status = 0;
3466                         goto bail;
3467                 }
3468 next:
3469                 offset += le16_to_cpu(de->rec_len);
3470                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3471         }
3472
3473 bail:
3474         brelse(bh);
3475         if (status)
3476                 mlog_errno(status);
3477
3478         return status;
3479 }
3480
3481 static int dx_leaf_sort_cmp(const void *a, const void *b)
3482 {
3483         const struct ocfs2_dx_entry *entry1 = a;
3484         const struct ocfs2_dx_entry *entry2 = b;
3485         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3486         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3487         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3488         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3489
3490         if (major_hash1 > major_hash2)
3491                 return 1;
3492         if (major_hash1 < major_hash2)
3493                 return -1;
3494
3495         /*
3496          * It is not strictly necessary to sort by minor
3497          */
3498         if (minor_hash1 > minor_hash2)
3499                 return 1;
3500         if (minor_hash1 < minor_hash2)
3501                 return -1;
3502         return 0;
3503 }
3504
3505 static void dx_leaf_sort_swap(void *a, void *b, int size)
3506 {
3507         struct ocfs2_dx_entry *entry1 = a;
3508         struct ocfs2_dx_entry *entry2 = b;
3509
3510         BUG_ON(size != sizeof(*entry1));
3511
3512         swap(*entry1, *entry2);
3513 }
3514
3515 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3516 {
3517         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3518         int i, num = le16_to_cpu(dl_list->de_num_used);
3519
3520         for (i = 0; i < (num - 1); i++) {
3521                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3522                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3523                         return 0;
3524         }
3525
3526         return 1;
3527 }
3528
3529 /*
3530  * Find the optimal value to split this leaf on. This expects the leaf
3531  * entries to be in sorted order.
3532  *
3533  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3534  * the hash we want to insert.
3535  *
3536  * This function is only concerned with the major hash - that which
3537  * determines which cluster an item belongs to.
3538  */
3539 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3540                                         u32 leaf_cpos, u32 insert_hash,
3541                                         u32 *split_hash)
3542 {
3543         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3544         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3545         int allsame;
3546
3547         /*
3548          * There's a couple rare, but nasty corner cases we have to
3549          * check for here. All of them involve a leaf where all value
3550          * have the same hash, which is what we look for first.
3551          *
3552          * Most of the time, all of the above is false, and we simply
3553          * pick the median value for a split.
3554          */
3555         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3556         if (allsame) {
3557                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3558
3559                 if (val == insert_hash) {
3560                         /*
3561                          * No matter where we would choose to split,
3562                          * the new entry would want to occupy the same
3563                          * block as these. Since there's no space left
3564                          * in their existing block, we know there
3565                          * won't be space after the split.
3566                          */
3567                         return -ENOSPC;
3568                 }
3569
3570                 if (val == leaf_cpos) {
3571                         /*
3572                          * Because val is the same as leaf_cpos (which
3573                          * is the smallest value this leaf can have),
3574                          * yet is not equal to insert_hash, then we
3575                          * know that insert_hash *must* be larger than
3576                          * val (and leaf_cpos). At least cpos+1 in value.
3577                          *
3578                          * We also know then, that there cannot be an
3579                          * adjacent extent (otherwise we'd be looking
3580                          * at it). Choosing this value gives us a
3581                          * chance to get some contiguousness.
3582                          */
3583                         *split_hash = leaf_cpos + 1;
3584                         return 0;
3585                 }
3586
3587                 if (val > insert_hash) {
3588                         /*
3589                          * val can not be the same as insert hash, and
3590                          * also must be larger than leaf_cpos. Also,
3591                          * we know that there can't be a leaf between
3592                          * cpos and val, otherwise the entries with
3593                          * hash 'val' would be there.
3594                          */
3595                         *split_hash = val;
3596                         return 0;
3597                 }
3598
3599                 *split_hash = insert_hash;
3600                 return 0;
3601         }
3602
3603         /*
3604          * Since the records are sorted and the checks above
3605          * guaranteed that not all records in this block are the same,
3606          * we simple travel forward, from the median, and pick the 1st
3607          * record whose value is larger than leaf_cpos.
3608          */
3609         for (i = (num_used / 2); i < num_used; i++)
3610                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3611                     leaf_cpos)
3612                         break;
3613
3614         BUG_ON(i == num_used); /* Should be impossible */
3615         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3616         return 0;
3617 }
3618
3619 /*
3620  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3621  * larger than split_hash into new_dx_leaves. We use a temporary
3622  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3623  *
3624  * Since the block offset inside a leaf (cluster) is a constant mask
3625  * of minor_hash, we can optimize - an item at block offset X within
3626  * the original cluster, will be at offset X within the new cluster.
3627  */
3628 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3629                                        handle_t *handle,
3630                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3631                                        struct buffer_head **orig_dx_leaves,
3632                                        struct buffer_head **new_dx_leaves,
3633                                        int num_dx_leaves)
3634 {
3635         int i, j, num_used;
3636         u32 major_hash;
3637         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3638         struct ocfs2_dx_entry_list *orig_list, *tmp_list;
3639         struct ocfs2_dx_entry *dx_entry;
3640
3641         tmp_list = &tmp_dx_leaf->dl_list;
3642
3643         for (i = 0; i < num_dx_leaves; i++) {
3644                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3645                 orig_list = &orig_dx_leaf->dl_list;
3646                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3647
3648                 num_used = le16_to_cpu(orig_list->de_num_used);
3649
3650                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3651                 tmp_list->de_num_used = cpu_to_le16(0);
3652                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3653
3654                 for (j = 0; j < num_used; j++) {
3655                         dx_entry = &orig_list->de_entries[j];
3656                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3657                         if (major_hash >= split_hash)
3658                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3659                                                               dx_entry);
3660                         else
3661                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3662                                                               dx_entry);
3663                 }
3664                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3665
3666                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3667                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3668         }
3669 }
3670
3671 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3672                                           struct ocfs2_dx_root_block *dx_root)
3673 {
3674         int credits = ocfs2_clusters_to_blocks(osb->sb, 3);
3675
3676         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3677         credits += ocfs2_quota_trans_credits(osb->sb);
3678         return credits;
3679 }
3680
3681 /*
3682  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3683  * half our entries into.
3684  */
3685 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3686                                   struct buffer_head *dx_root_bh,
3687                                   struct buffer_head *dx_leaf_bh,
3688                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3689                                   u64 leaf_blkno)
3690 {
3691         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3692         int credits, ret, i, num_used, did_quota = 0;
3693         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3694         u64 orig_leaves_start;
3695         int num_dx_leaves;
3696         struct buffer_head **orig_dx_leaves = NULL;
3697         struct buffer_head **new_dx_leaves = NULL;
3698         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3699         struct ocfs2_extent_tree et;
3700         handle_t *handle = NULL;
3701         struct ocfs2_dx_root_block *dx_root;
3702         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3703
3704         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3705                                      (unsigned long long)leaf_blkno,
3706                                      insert_hash);
3707
3708         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3709
3710         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3711         /*
3712          * XXX: This is a rather large limit. We should use a more
3713          * realistic value.
3714          */
3715         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3716                 return -ENOSPC;
3717
3718         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3719         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3720                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3721                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3722                      (unsigned long long)leaf_blkno, num_used);
3723                 ret = -EIO;
3724                 goto out;
3725         }
3726
3727         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3728         if (!orig_dx_leaves) {
3729                 ret = -ENOMEM;
3730                 mlog_errno(ret);
3731                 goto out;
3732         }
3733
3734         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3735         if (!new_dx_leaves) {
3736                 ret = -ENOMEM;
3737                 mlog_errno(ret);
3738                 goto out;
3739         }
3740
3741         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3742         if (ret) {
3743                 if (ret != -ENOSPC)
3744                         mlog_errno(ret);
3745                 goto out;
3746         }
3747
3748         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3749         handle = ocfs2_start_trans(osb, credits);
3750         if (IS_ERR(handle)) {
3751                 ret = PTR_ERR(handle);
3752                 handle = NULL;
3753                 mlog_errno(ret);
3754                 goto out;
3755         }
3756
3757         ret = dquot_alloc_space_nodirty(dir,
3758                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3759         if (ret)
3760                 goto out_commit;
3761         did_quota = 1;
3762
3763         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3764                                       OCFS2_JOURNAL_ACCESS_WRITE);
3765         if (ret) {
3766                 mlog_errno(ret);
3767                 goto out_commit;
3768         }
3769
3770         /*
3771          * This block is changing anyway, so we can sort it in place.
3772          */
3773         sort(dx_leaf->dl_list.de_entries, num_used,
3774              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3775              dx_leaf_sort_swap);
3776
3777         ocfs2_journal_dirty(handle, dx_leaf_bh);
3778
3779         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3780                                            &split_hash);
3781         if (ret) {
3782                 mlog_errno(ret);
3783                 goto  out_commit;
3784         }
3785
3786         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3787
3788         /*
3789          * We have to carefully order operations here. There are items
3790          * which want to be in the new cluster before insert, but in
3791          * order to put those items in the new cluster, we alter the
3792          * old cluster. A failure to insert gets nasty.
3793          *
3794          * So, start by reserving writes to the old
3795          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3796          * the new cluster for us, before inserting it. The insert
3797          * won't happen if there's an error before that. Once the
3798          * insert is done then, we can transfer from one leaf into the
3799          * other without fear of hitting any error.
3800          */
3801
3802         /*
3803          * The leaf transfer wants some scratch space so that we don't
3804          * wind up doing a bunch of expensive memmove().
3805          */
3806         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3807         if (!tmp_dx_leaf) {
3808                 ret = -ENOMEM;
3809                 mlog_errno(ret);
3810                 goto out_commit;
3811         }
3812
3813         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3814         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3815                                    orig_dx_leaves);
3816         if (ret) {
3817                 mlog_errno(ret);
3818                 goto out_commit;
3819         }
3820
3821         cpos = split_hash;
3822         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3823                                        data_ac, meta_ac, new_dx_leaves,
3824                                        num_dx_leaves);
3825         if (ret) {
3826                 mlog_errno(ret);
3827                 goto out_commit;
3828         }
3829
3830         for (i = 0; i < num_dx_leaves; i++) {
3831                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3832                                               orig_dx_leaves[i],
3833                                               OCFS2_JOURNAL_ACCESS_WRITE);
3834                 if (ret) {
3835                         mlog_errno(ret);
3836                         goto out_commit;
3837                 }
3838
3839                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3840                                               new_dx_leaves[i],
3841                                               OCFS2_JOURNAL_ACCESS_WRITE);
3842                 if (ret) {
3843                         mlog_errno(ret);
3844                         goto out_commit;
3845                 }
3846         }
3847
3848         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3849                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3850
3851 out_commit:
3852         if (ret < 0 && did_quota)
3853                 dquot_free_space_nodirty(dir,
3854                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3855
3856         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3857         ocfs2_commit_trans(osb, handle);
3858
3859 out:
3860         if (orig_dx_leaves || new_dx_leaves) {
3861                 for (i = 0; i < num_dx_leaves; i++) {
3862                         if (orig_dx_leaves)
3863                                 brelse(orig_dx_leaves[i]);
3864                         if (new_dx_leaves)
3865                                 brelse(new_dx_leaves[i]);
3866                 }
3867                 kfree(orig_dx_leaves);
3868                 kfree(new_dx_leaves);
3869         }
3870
3871         if (meta_ac)
3872                 ocfs2_free_alloc_context(meta_ac);
3873         if (data_ac)
3874                 ocfs2_free_alloc_context(data_ac);
3875
3876         kfree(tmp_dx_leaf);
3877         return ret;
3878 }
3879
3880 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3881                                    struct buffer_head *di_bh,
3882                                    struct buffer_head *dx_root_bh,
3883                                    const char *name, int namelen,
3884                                    struct ocfs2_dir_lookup_result *lookup)
3885 {
3886         int ret, rebalanced = 0;
3887         struct ocfs2_dx_root_block *dx_root;
3888         struct buffer_head *dx_leaf_bh = NULL;
3889         struct ocfs2_dx_leaf *dx_leaf;
3890         u64 blkno;
3891         u32 leaf_cpos;
3892
3893         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3894
3895 restart_search:
3896         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3897                                   &leaf_cpos, &blkno);
3898         if (ret) {
3899                 mlog_errno(ret);
3900                 goto out;
3901         }
3902
3903         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3904         if (ret) {
3905                 mlog_errno(ret);
3906                 goto out;
3907         }
3908
3909         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3910
3911         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3912             le16_to_cpu(dx_leaf->dl_list.de_count)) {
3913                 if (rebalanced) {
3914                         /*
3915                          * Rebalancing should have provided us with
3916                          * space in an appropriate leaf.
3917                          *
3918                          * XXX: Is this an abnormal condition then?
3919                          * Should we print a message here?
3920                          */
3921                         ret = -ENOSPC;
3922                         goto out;
3923                 }
3924
3925                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3926                                              &lookup->dl_hinfo, leaf_cpos,
3927                                              blkno);
3928                 if (ret) {
3929                         if (ret != -ENOSPC)
3930                                 mlog_errno(ret);
3931                         goto out;
3932                 }
3933
3934                 /*
3935                  * Restart the lookup. The rebalance might have
3936                  * changed which block our item fits into. Mark our
3937                  * progress, so we only execute this once.
3938                  */
3939                 brelse(dx_leaf_bh);
3940                 dx_leaf_bh = NULL;
3941                 rebalanced = 1;
3942                 goto restart_search;
3943         }
3944
3945         lookup->dl_dx_leaf_bh = dx_leaf_bh;
3946         dx_leaf_bh = NULL;
3947
3948 out:
3949         brelse(dx_leaf_bh);
3950         return ret;
3951 }
3952
3953 static int ocfs2_search_dx_free_list(struct inode *dir,
3954                                      struct buffer_head *dx_root_bh,
3955                                      int namelen,
3956                                      struct ocfs2_dir_lookup_result *lookup)
3957 {
3958         int ret = -ENOSPC;
3959         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
3960         struct ocfs2_dir_block_trailer *db;
3961         u64 next_block;
3962         int rec_len = OCFS2_DIR_REC_LEN(namelen);
3963         struct ocfs2_dx_root_block *dx_root;
3964
3965         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3966         next_block = le64_to_cpu(dx_root->dr_free_blk);
3967
3968         while (next_block) {
3969                 brelse(prev_leaf_bh);
3970                 prev_leaf_bh = leaf_bh;
3971                 leaf_bh = NULL;
3972
3973                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
3974                 if (ret) {
3975                         mlog_errno(ret);
3976                         goto out;
3977                 }
3978
3979                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
3980                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
3981                         lookup->dl_leaf_bh = leaf_bh;
3982                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
3983                         leaf_bh = NULL;
3984                         prev_leaf_bh = NULL;
3985                         break;
3986                 }
3987
3988                 next_block = le64_to_cpu(db->db_free_next);
3989         }
3990
3991         if (!next_block)
3992                 ret = -ENOSPC;
3993
3994 out:
3995
3996         brelse(leaf_bh);
3997         brelse(prev_leaf_bh);
3998         return ret;
3999 }
4000
4001 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4002                                        struct buffer_head *dx_root_bh)
4003 {
4004         int ret, num_dx_leaves, i, j, did_quota = 0;
4005         struct buffer_head **dx_leaves = NULL;
4006         struct ocfs2_extent_tree et;
4007         u64 insert_blkno;
4008         struct ocfs2_alloc_context *data_ac = NULL;
4009         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4010         handle_t *handle = NULL;
4011         struct ocfs2_dx_root_block *dx_root;
4012         struct ocfs2_dx_entry_list *entry_list;
4013         struct ocfs2_dx_entry *dx_entry;
4014         struct ocfs2_dx_leaf *target_leaf;
4015
4016         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4017         if (ret) {
4018                 mlog_errno(ret);
4019                 goto out;
4020         }
4021
4022         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4023         if (!dx_leaves) {
4024                 ret = -ENOMEM;
4025                 mlog_errno(ret);
4026                 goto out;
4027         }
4028
4029         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4030         if (IS_ERR(handle)) {
4031                 ret = PTR_ERR(handle);
4032                 mlog_errno(ret);
4033                 goto out;
4034         }
4035
4036         ret = dquot_alloc_space_nodirty(dir,
4037                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4038         if (ret)
4039                 goto out_commit;
4040         did_quota = 1;
4041
4042         /*
4043          * We do this up front, before the allocation, so that a
4044          * failure to add the dx_root_bh to the journal won't result
4045          * us losing clusters.
4046          */
4047         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4048                                       OCFS2_JOURNAL_ACCESS_WRITE);
4049         if (ret) {
4050                 mlog_errno(ret);
4051                 goto out_commit;
4052         }
4053
4054         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4055                                          num_dx_leaves, &insert_blkno);
4056         if (ret) {
4057                 mlog_errno(ret);
4058                 goto out_commit;
4059         }
4060
4061         /*
4062          * Transfer the entries from our dx_root into the appropriate
4063          * block
4064          */
4065         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4066         entry_list = &dx_root->dr_entries;
4067
4068         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4069                 dx_entry = &entry_list->de_entries[i];
4070
4071                 j = __ocfs2_dx_dir_hash_idx(osb,
4072                                             le32_to_cpu(dx_entry->dx_minor_hash));
4073                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4074
4075                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4076
4077                 /* Each leaf has been passed to the journal already
4078                  * via __ocfs2_dx_dir_new_cluster() */
4079         }
4080
4081         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4082         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4083                offsetof(struct ocfs2_dx_root_block, dr_list));
4084         dx_root->dr_list.l_count =
4085                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4086
4087         /* This should never fail considering we start with an empty
4088          * dx_root. */
4089         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4090         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4091         if (ret)
4092                 mlog_errno(ret);
4093         did_quota = 0;
4094
4095         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4096         ocfs2_journal_dirty(handle, dx_root_bh);
4097
4098 out_commit:
4099         if (ret < 0 && did_quota)
4100                 dquot_free_space_nodirty(dir,
4101                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4102
4103         ocfs2_commit_trans(osb, handle);
4104
4105 out:
4106         if (data_ac)
4107                 ocfs2_free_alloc_context(data_ac);
4108
4109         if (dx_leaves) {
4110                 for (i = 0; i < num_dx_leaves; i++)
4111                         brelse(dx_leaves[i]);
4112                 kfree(dx_leaves);
4113         }
4114         return ret;
4115 }
4116
4117 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4118 {
4119         struct ocfs2_dx_root_block *dx_root;
4120         struct ocfs2_dx_entry_list *entry_list;
4121
4122         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4123         entry_list = &dx_root->dr_entries;
4124
4125         if (le16_to_cpu(entry_list->de_num_used) >=
4126             le16_to_cpu(entry_list->de_count))
4127                 return -ENOSPC;
4128
4129         return 0;
4130 }
4131
4132 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4133                                            struct buffer_head *di_bh,
4134                                            const char *name,
4135                                            int namelen,
4136                                            struct ocfs2_dir_lookup_result *lookup)
4137 {
4138         int ret, free_dx_root = 1;
4139         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4140         struct buffer_head *dx_root_bh = NULL;
4141         struct buffer_head *leaf_bh = NULL;
4142         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4143         struct ocfs2_dx_root_block *dx_root;
4144
4145         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4146         if (ret) {
4147                 mlog_errno(ret);
4148                 goto out;
4149         }
4150
4151         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4152         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4153                 ret = -ENOSPC;
4154                 mlog_errno(ret);
4155                 goto out;
4156         }
4157
4158         if (ocfs2_dx_root_inline(dx_root)) {
4159                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4160
4161                 if (ret == 0)
4162                         goto search_el;
4163
4164                 /*
4165                  * We ran out of room in the root block. Expand it to
4166                  * an extent, then allow ocfs2_find_dir_space_dx to do
4167                  * the rest.
4168                  */
4169                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4170                 if (ret) {
4171                         mlog_errno(ret);
4172                         goto out;
4173                 }
4174         }
4175
4176         /*
4177          * Insert preparation for an indexed directory is split into two
4178          * steps. The call to find_dir_space_dx reserves room in the index for
4179          * an additional item. If we run out of space there, it's a real error
4180          * we can't continue on.
4181          */
4182         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4183                                       namelen, lookup);
4184         if (ret) {
4185                 mlog_errno(ret);
4186                 goto out;
4187         }
4188
4189 search_el:
4190         /*
4191          * Next, we need to find space in the unindexed tree. This call
4192          * searches using the free space linked list. If the unindexed tree
4193          * lacks sufficient space, we'll expand it below. The expansion code
4194          * is smart enough to add any new blocks to the free space list.
4195          */
4196         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4197         if (ret && ret != -ENOSPC) {
4198                 mlog_errno(ret);
4199                 goto out;
4200         }
4201
4202         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4203         lookup->dl_dx_root_bh = dx_root_bh;
4204         free_dx_root = 0;
4205
4206         if (ret == -ENOSPC) {
4207                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4208
4209                 if (ret) {
4210                         mlog_errno(ret);
4211                         goto out;
4212                 }
4213
4214                 /*
4215                  * We make the assumption here that new leaf blocks are added
4216                  * to the front of our free list.
4217                  */
4218                 lookup->dl_prev_leaf_bh = NULL;
4219                 lookup->dl_leaf_bh = leaf_bh;
4220         }
4221
4222 out:
4223         if (free_dx_root)
4224                 brelse(dx_root_bh);
4225         return ret;
4226 }
4227
4228 /*
4229  * Get a directory ready for insert. Any directory allocation required
4230  * happens here. Success returns zero, and enough context in the dir
4231  * lookup result that ocfs2_add_entry() will be able complete the task
4232  * with minimal performance impact.
4233  */
4234 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4235                                  struct inode *dir,
4236                                  struct buffer_head *parent_fe_bh,
4237                                  const char *name,
4238                                  int namelen,
4239                                  struct ocfs2_dir_lookup_result *lookup)
4240 {
4241         int ret;
4242         unsigned int blocks_wanted = 1;
4243         struct buffer_head *bh = NULL;
4244
4245         trace_ocfs2_prepare_dir_for_insert(
4246                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4247
4248         if (!namelen) {
4249                 ret = -EINVAL;
4250                 mlog_errno(ret);
4251                 goto out;
4252         }
4253
4254         /*
4255          * Do this up front to reduce confusion.
4256          *
4257          * The directory might start inline, then be turned into an
4258          * indexed one, in which case we'd need to hash deep inside
4259          * ocfs2_find_dir_space_id(). Since
4260          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4261          * done, there seems no point in spreading out the calls. We
4262          * can optimize away the case where the file system doesn't
4263          * support indexing.
4264          */
4265         if (ocfs2_supports_indexed_dirs(osb))
4266                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4267
4268         if (ocfs2_dir_indexed(dir)) {
4269                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4270                                                       name, namelen, lookup);
4271                 if (ret)
4272                         mlog_errno(ret);
4273                 goto out;
4274         }
4275
4276         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4277                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4278                                               namelen, &bh, &blocks_wanted);
4279         } else
4280                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4281
4282         if (ret && ret != -ENOSPC) {
4283                 mlog_errno(ret);
4284                 goto out;
4285         }
4286
4287         if (ret == -ENOSPC) {
4288                 /*
4289                  * We have to expand the directory to add this name.
4290                  */
4291                 BUG_ON(bh);
4292
4293                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4294                                        lookup, &bh);
4295                 if (ret) {
4296                         if (ret != -ENOSPC)
4297                                 mlog_errno(ret);
4298                         goto out;
4299                 }
4300
4301                 BUG_ON(!bh);
4302         }
4303
4304         lookup->dl_leaf_bh = bh;
4305         bh = NULL;
4306 out:
4307         brelse(bh);
4308         return ret;
4309 }
4310
4311 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4312                                      struct buffer_head *di_bh,
4313                                      struct buffer_head *dx_root_bh)
4314 {
4315         int ret;
4316         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4317         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4318         struct ocfs2_dx_root_block *dx_root;
4319         struct inode *dx_alloc_inode = NULL;
4320         struct buffer_head *dx_alloc_bh = NULL;
4321         handle_t *handle;
4322         u64 blk;
4323         u16 bit;
4324         u64 bg_blkno;
4325
4326         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4327
4328         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4329                                         EXTENT_ALLOC_SYSTEM_INODE,
4330                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4331         if (!dx_alloc_inode) {
4332                 ret = -ENOMEM;
4333                 mlog_errno(ret);
4334                 goto out;
4335         }
4336         inode_lock(dx_alloc_inode);
4337
4338         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4339         if (ret) {
4340                 mlog_errno(ret);
4341                 goto out_mutex;
4342         }
4343
4344         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4345         if (IS_ERR(handle)) {
4346                 ret = PTR_ERR(handle);
4347                 mlog_errno(ret);
4348                 goto out_unlock;
4349         }
4350
4351         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4352                                       OCFS2_JOURNAL_ACCESS_WRITE);
4353         if (ret) {
4354                 mlog_errno(ret);
4355                 goto out_commit;
4356         }
4357
4358         spin_lock(&OCFS2_I(dir)->ip_lock);
4359         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4360         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4361         spin_unlock(&OCFS2_I(dir)->ip_lock);
4362         di->i_dx_root = cpu_to_le64(0ULL);
4363         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4364
4365         ocfs2_journal_dirty(handle, di_bh);
4366
4367         blk = le64_to_cpu(dx_root->dr_blkno);
4368         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4369         if (dx_root->dr_suballoc_loc)
4370                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4371         else
4372                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4373         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4374                                        bit, bg_blkno, 1);
4375         if (ret)
4376                 mlog_errno(ret);
4377
4378 out_commit:
4379         ocfs2_commit_trans(osb, handle);
4380
4381 out_unlock:
4382         ocfs2_inode_unlock(dx_alloc_inode, 1);
4383
4384 out_mutex:
4385         inode_unlock(dx_alloc_inode);
4386         brelse(dx_alloc_bh);
4387 out:
4388         iput(dx_alloc_inode);
4389         return ret;
4390 }
4391
4392 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4393 {
4394         int ret;
4395         unsigned int clen;
4396         u32 major_hash = UINT_MAX, p_cpos, cpos;
4397         u64 blkno;
4398         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4399         struct buffer_head *dx_root_bh = NULL;
4400         struct ocfs2_dx_root_block *dx_root;
4401         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4402         struct ocfs2_cached_dealloc_ctxt dealloc;
4403         struct ocfs2_extent_tree et;
4404
4405         ocfs2_init_dealloc_ctxt(&dealloc);
4406
4407         if (!ocfs2_dir_indexed(dir))
4408                 return 0;
4409
4410         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4411         if (ret) {
4412                 mlog_errno(ret);
4413                 goto out;
4414         }
4415         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4416
4417         if (ocfs2_dx_root_inline(dx_root))
4418                 goto remove_index;
4419
4420         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4421
4422         /* XXX: What if dr_clusters is too large? */
4423         while (le32_to_cpu(dx_root->dr_clusters)) {
4424                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4425                                               major_hash, &cpos, &blkno, &clen);
4426                 if (ret) {
4427                         mlog_errno(ret);
4428                         goto out;
4429                 }
4430
4431                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4432
4433                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4434                                                &dealloc, 0, false);
4435                 if (ret) {
4436                         mlog_errno(ret);
4437                         goto out;
4438                 }
4439
4440                 if (cpos == 0)
4441                         break;
4442
4443                 major_hash = cpos - 1;
4444         }
4445
4446 remove_index:
4447         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4448         if (ret) {
4449                 mlog_errno(ret);
4450                 goto out;
4451         }
4452
4453         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4454 out:
4455         ocfs2_schedule_truncate_log_flush(osb, 1);
4456         ocfs2_run_deallocs(osb, &dealloc);
4457
4458         brelse(dx_root_bh);
4459         return ret;
4460 }