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 ([email protected])
  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                                  char *buf,
 301                                  unsigned int size,
 302                                  unsigned long offset)
 303 {
 304         const char *error_msg = NULL;
 305         const int rlen = le16_to_cpu(de->rec_len);
 306         const unsigned long next_offset = ((char *) de - buf) + rlen;
 307
 308         if (unlikely(rlen < OCFS2_DIR_REC_LEN(1)))
 309                 error_msg = "rec_len is smaller than minimal";
 310         else if (unlikely(rlen % 4 != 0))
 311                 error_msg = "rec_len % 4 != 0";
 312         else if (unlikely(rlen < OCFS2_DIR_REC_LEN(de->name_len)))
 313                 error_msg = "rec_len is too small for name_len";
 314         else if (unlikely(next_offset > size))
 315                 error_msg = "directory entry overrun";
 316         else if (unlikely(next_offset > size - OCFS2_DIR_REC_LEN(1)) &&
 317                  next_offset != size)
 318                 error_msg = "directory entry too close to end";
 319
 320         if (unlikely(error_msg != NULL))
 321                 mlog(ML_ERROR, "bad entry in directory #%llu: %s - "
 322                      "offset=%lu, inode=%llu, rec_len=%d, name_len=%d\n",
 323                      (unsigned long long)OCFS2_I(dir)->ip_blkno, error_msg,
 324                      offset, (unsigned long long)le64_to_cpu(de->inode), rlen,
 325                      de->name_len);
 326
 327         return error_msg == NULL ? 1 : 0;
 328 }
 329
 330 static inline int ocfs2_match(int len,
 331                               const char * const name,
 332                               struct ocfs2_dir_entry *de)
 333 {
 334         if (len != de->name_len)
 335                 return 0;
 336         if (!de->inode)
 337                 return 0;
 338         return !memcmp(name, de->name, len);
 339 }
 340
 341 /*
 342  * Returns 0 if not found, -1 on failure, and 1 on success
 343  */
 344 static inline int ocfs2_search_dirblock(struct buffer_head *bh,
 345                                         struct inode *dir,
 346                                         const char *name, int namelen,
 347                                         unsigned long offset,
 348                                         char *first_de,
 349                                         unsigned int bytes,
 350                                         struct ocfs2_dir_entry **res_dir)
 351 {
 352         struct ocfs2_dir_entry *de;
 353         char *dlimit, *de_buf;
 354         int de_len;
 355         int ret = 0;
 356
 357         de_buf = first_de;
 358         dlimit = de_buf + bytes;
 359
 360         while (de_buf < dlimit - OCFS2_DIR_MEMBER_LEN) {
 361                 /* this code is executed quadratically often */
 362                 /* do minimal checking `by hand' */
 363
 364                 de = (struct ocfs2_dir_entry *) de_buf;
 365
 366                 if (de->name + namelen <= dlimit &&
 367                     ocfs2_match(namelen, name, de)) {
 368                         /* found a match - just to be sure, do a full check */
 369                         if (!ocfs2_check_dir_entry(dir, de, bh, first_de,
 370                                                    bytes, offset)) {
 371                                 ret = -1;
 372                                 goto bail;
 373                         }
 374                         *res_dir = de;
 375                         ret = 1;
 376                         goto bail;
 377                 }
 378
 379                 /* prevent looping on a bad block */
 380                 de_len = le16_to_cpu(de->rec_len);
 381                 if (de_len <= 0) {
 382                         ret = -1;
 383                         goto bail;
 384                 }
 385
 386                 de_buf += de_len;
 387                 offset += de_len;
 388         }
 389
 390 bail:
 391         trace_ocfs2_search_dirblock(ret);
 392         return ret;
 393 }
 394
 395 static struct buffer_head *ocfs2_find_entry_id(const char *name,
 396                                                int namelen,
 397                                                struct inode *dir,
 398                                                struct ocfs2_dir_entry **res_dir)
 399 {
 400         int ret, found;
 401         struct buffer_head *di_bh = NULL;
 402         struct ocfs2_dinode *di;
 403         struct ocfs2_inline_data *data;
 404
 405         ret = ocfs2_read_inode_block(dir, &di_bh);
 406         if (ret) {
 407                 mlog_errno(ret);
 408                 goto out;
 409         }
 410
 411         di = (struct ocfs2_dinode *)di_bh->b_data;
 412         data = &di->id2.i_data;
 413
 414         found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
 415                                       data->id_data, i_size_read(dir), res_dir);
 416         if (found == 1)
 417                 return di_bh;
 418
 419         brelse(di_bh);
 420 out:
 421         return NULL;
 422 }
 423
 424 static int ocfs2_validate_dir_block(struct super_block *sb,
 425                                     struct buffer_head *bh)
 426 {
 427         int rc;
 428         struct ocfs2_dir_block_trailer *trailer =
 429                 ocfs2_trailer_from_bh(bh, sb);
 430
 431
 432         /*
 433          * We don't validate dirents here, that's handled
 434          * in-place when the code walks them.
 435          */
 436         trace_ocfs2_validate_dir_block((unsigned long long)bh->b_blocknr);
 437
 438         BUG_ON(!buffer_uptodate(bh));
 439
 440         /*
 441          * If the ecc fails, we return the error but otherwise
 442          * leave the filesystem running.  We know any error is
 443          * local to this block.
 444          *
 445          * Note that we are safe to call this even if the directory
 446          * doesn't have a trailer.  Filesystems without metaecc will do
 447          * nothing, and filesystems with it will have one.
 448          */
 449         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
 450         if (rc)
 451                 mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
 452                      (unsigned long long)bh->b_blocknr);
 453
 454         return rc;
 455 }
 456
 457 /*
 458  * Validate a directory trailer.
 459  *
 460  * We check the trailer here rather than in ocfs2_validate_dir_block()
 461  * because that function doesn't have the inode to test.
 462  */
 463 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
 464 {
 465         int rc = 0;
 466         struct ocfs2_dir_block_trailer *trailer;
 467
 468         trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
 469         if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
 470                 rc = ocfs2_error(dir->i_sb,
 471                                  "Invalid dirblock #%llu: signature = %.*s\n",
 472                                  (unsigned long long)bh->b_blocknr, 7,
 473                                  trailer->db_signature);
 474                 goto out;
 475         }
 476         if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
 477                 rc = ocfs2_error(dir->i_sb,
 478                                  "Directory block #%llu has an invalid db_blkno of %llu\n",
 479                                  (unsigned long long)bh->b_blocknr,
 480                                  (unsigned long long)le64_to_cpu(trailer->db_blkno));
 481                 goto out;
 482         }
 483         if (le64_to_cpu(trailer->db_parent_dinode) !=
 484             OCFS2_I(dir)->ip_blkno) {
 485                 rc = ocfs2_error(dir->i_sb,
 486                                  "Directory block #%llu on dinode #%llu has an invalid parent_dinode of %llu\n",
 487                                  (unsigned long long)bh->b_blocknr,
 488                                  (unsigned long long)OCFS2_I(dir)->ip_blkno,
 489                                  (unsigned long long)le64_to_cpu(trailer->db_blkno));
 490                 goto out;
 491         }
 492 out:
 493         return rc;
 494 }
 495
 496 /*
 497  * This function forces all errors to -EIO for consistency with its
 498  * predecessor, ocfs2_bread().  We haven't audited what returning the
 499  * real error codes would do to callers.  We log the real codes with
 500  * mlog_errno() before we squash them.
 501  */
 502 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
 503                                 struct buffer_head **bh, int flags)
 504 {
 505         int rc = 0;
 506         struct buffer_head *tmp = *bh;
 507
 508         rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
 509                                     ocfs2_validate_dir_block);
 510         if (rc) {
 511                 mlog_errno(rc);
 512                 goto out;
 513         }
 514
 515         if (!(flags & OCFS2_BH_READAHEAD) &&
 516             ocfs2_supports_dir_trailer(inode)) {
 517                 rc = ocfs2_check_dir_trailer(inode, tmp);
 518                 if (rc) {
 519                         if (!*bh)
 520                                 brelse(tmp);
 521                         mlog_errno(rc);
 522                         goto out;
 523                 }
 524         }
 525
 526         /* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
 527         if (!*bh)
 528                 *bh = tmp;
 529
 530 out:
 531         return rc ? -EIO : 0;
 532 }
 533
 534 /*
 535  * Read the block at 'phys' which belongs to this directory
 536  * inode. This function does no virtual->physical block translation -
 537  * what's passed in is assumed to be a valid directory block.
 538  */
 539 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
 540                                        struct buffer_head **bh)
 541 {
 542         int ret;
 543         struct buffer_head *tmp = *bh;
 544
 545         ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
 546                                ocfs2_validate_dir_block);
 547         if (ret) {
 548                 mlog_errno(ret);
 549                 goto out;
 550         }
 551
 552         if (ocfs2_supports_dir_trailer(dir)) {
 553                 ret = ocfs2_check_dir_trailer(dir, tmp);
 554                 if (ret) {
 555                         if (!*bh)
 556                                 brelse(tmp);
 557                         mlog_errno(ret);
 558                         goto out;
 559                 }
 560         }
 561
 562         if (!ret && !*bh)
 563                 *bh = tmp;
 564 out:
 565         return ret;
 566 }
 567
 568 static int ocfs2_validate_dx_root(struct super_block *sb,
 569                                   struct buffer_head *bh)
 570 {
 571         int ret;
 572         struct ocfs2_dx_root_block *dx_root;
 573
 574         BUG_ON(!buffer_uptodate(bh));
 575
 576         dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
 577
 578         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
 579         if (ret) {
 580                 mlog(ML_ERROR,
 581                      "Checksum failed for dir index root block %llu\n",
 582                      (unsigned long long)bh->b_blocknr);
 583                 return ret;
 584         }
 585
 586         if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
 587                 ret = ocfs2_error(sb,
 588                                   "Dir Index Root # %llu has bad signature %.*s\n",
 589                                   (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
 590                                   7, dx_root->dr_signature);
 591         }
 592
 593         return ret;
 594 }
 595
 596 static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
 597                               struct buffer_head **dx_root_bh)
 598 {
 599         int ret;
 600         u64 blkno = le64_to_cpu(di->i_dx_root);
 601         struct buffer_head *tmp = *dx_root_bh;
 602
 603         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 604                                ocfs2_validate_dx_root);
 605
 606         /* If ocfs2_read_block() got us a new bh, pass it up. */
 607         if (!ret && !*dx_root_bh)
 608                 *dx_root_bh = tmp;
 609
 610         return ret;
 611 }
 612
 613 static int ocfs2_validate_dx_leaf(struct super_block *sb,
 614                                   struct buffer_head *bh)
 615 {
 616         int ret;
 617         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
 618
 619         BUG_ON(!buffer_uptodate(bh));
 620
 621         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
 622         if (ret) {
 623                 mlog(ML_ERROR,
 624                      "Checksum failed for dir index leaf block %llu\n",
 625                      (unsigned long long)bh->b_blocknr);
 626                 return ret;
 627         }
 628
 629         if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
 630                 ret = ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s\n",
 631                                   7, dx_leaf->dl_signature);
 632         }
 633
 634         return ret;
 635 }
 636
 637 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
 638                               struct buffer_head **dx_leaf_bh)
 639 {
 640         int ret;
 641         struct buffer_head *tmp = *dx_leaf_bh;
 642
 643         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
 644                                ocfs2_validate_dx_leaf);
 645
 646         /* If ocfs2_read_block() got us a new bh, pass it up. */
 647         if (!ret && !*dx_leaf_bh)
 648                 *dx_leaf_bh = tmp;
 649
 650         return ret;
 651 }
 652
 653 /*
 654  * Read a series of dx_leaf blocks. This expects all buffer_head
 655  * pointers to be NULL on function entry.
 656  */
 657 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
 658                                 struct buffer_head **dx_leaf_bhs)
 659 {
 660         int ret;
 661
 662         ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
 663                                 ocfs2_validate_dx_leaf);
 664         if (ret)
 665                 mlog_errno(ret);
 666
 667         return ret;
 668 }
 669
 670 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
 671                                                struct inode *dir,
 672                                                struct ocfs2_dir_entry **res_dir)
 673 {
 674         struct super_block *sb;
 675         struct buffer_head *bh_use[NAMEI_RA_SIZE];
 676         struct buffer_head *bh, *ret = NULL;
 677         unsigned long start, block, b;
 678         int ra_max = 0;         /* Number of bh's in the readahead
 679                                    buffer, bh_use[] */
 680         int ra_ptr = 0;         /* Current index into readahead
 681                                    buffer */
 682         int num = 0;
 683         int nblocks, i;
 684
 685         sb = dir->i_sb;
 686
 687         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 688         start = OCFS2_I(dir)->ip_dir_start_lookup;
 689         if (start >= nblocks)
 690                 start = 0;
 691         block = start;
 692
 693 restart:
 694         do {
 695                 /*
 696                  * We deal with the read-ahead logic here.
 697                  */
 698                 if (ra_ptr >= ra_max) {
 699                         /* Refill the readahead buffer */
 700                         ra_ptr = 0;
 701                         b = block;
 702                         for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
 703                                 /*
 704                                  * Terminate if we reach the end of the
 705                                  * directory and must wrap, or if our
 706                                  * search has finished at this block.
 707                                  */
 708                                 if (b >= nblocks || (num && block == start)) {
 709                                         bh_use[ra_max] = NULL;
 710                                         break;
 711                                 }
 712                                 num++;
 713
 714                                 bh = NULL;
 715                                 ocfs2_read_dir_block(dir, b++, &bh,
 716                                                            OCFS2_BH_READAHEAD);
 717                                 bh_use[ra_max] = bh;
 718                         }
 719                 }
 720                 if ((bh = bh_use[ra_ptr++]) == NULL)
 721                         goto next;
 722                 if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
 723                         /* read error, skip block & hope for the best.
 724                          * ocfs2_read_dir_block() has released the bh. */
 725                         mlog(ML_ERROR, "reading directory %llu, "
 726                                     "offset %lu\n",
 727                                     (unsigned long long)OCFS2_I(dir)->ip_blkno,
 728                                     block);
 729                         goto next;
 730                 }
 731                 i = ocfs2_search_dirblock(bh, dir, name, namelen,
 732                                           block << sb->s_blocksize_bits,
 733                                           bh->b_data, sb->s_blocksize,
 734                                           res_dir);
 735                 if (i == 1) {
 736                         OCFS2_I(dir)->ip_dir_start_lookup = block;
 737                         ret = bh;
 738                         goto cleanup_and_exit;
 739                 } else {
 740                         brelse(bh);
 741                         if (i < 0)
 742                                 goto cleanup_and_exit;
 743                 }
 744         next:
 745                 if (++block >= nblocks)
 746                         block = 0;
 747         } while (block != start);
 748
 749         /*
 750          * If the directory has grown while we were searching, then
 751          * search the last part of the directory before giving up.
 752          */
 753         block = nblocks;
 754         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
 755         if (block < nblocks) {
 756                 start = 0;
 757                 goto restart;
 758         }
 759
 760 cleanup_and_exit:
 761         /* Clean up the read-ahead blocks */
 762         for (; ra_ptr < ra_max; ra_ptr++)
 763                 brelse(bh_use[ra_ptr]);
 764
 765         trace_ocfs2_find_entry_el(ret);
 766         return ret;
 767 }
 768
 769 static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
 770                                    struct ocfs2_extent_list *el,
 771                                    u32 major_hash,
 772                                    u32 *ret_cpos,
 773                                    u64 *ret_phys_blkno,
 774                                    unsigned int *ret_clen)
 775 {
 776         int ret = 0, i, found;
 777         struct buffer_head *eb_bh = NULL;
 778         struct ocfs2_extent_block *eb;
 779         struct ocfs2_extent_rec *rec = NULL;
 780
 781         if (el->l_tree_depth) {
 782                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
 783                                       &eb_bh);
 784                 if (ret) {
 785                         mlog_errno(ret);
 786                         goto out;
 787                 }
 788
 789                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 790                 el = &eb->h_list;
 791
 792                 if (el->l_tree_depth) {
 793                         ret = ocfs2_error(inode->i_sb,
 794                                           "Inode %lu has non zero tree depth in btree tree block %llu\n",
 795                                           inode->i_ino,
 796                                           (unsigned long long)eb_bh->b_blocknr);
 797                         goto out;
 798                 }
 799         }
 800
 801         found = 0;
 802         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
 803                 rec = &el->l_recs[i];
 804
 805                 if (le32_to_cpu(rec->e_cpos) <= major_hash) {
 806                         found = 1;
 807                         break;
 808                 }
 809         }
 810
 811         if (!found) {
 812                 ret = ocfs2_error(inode->i_sb,
 813                                   "Inode %lu has bad extent record (%u, %u, 0) in btree\n",
 814                                   inode->i_ino,
 815                                   le32_to_cpu(rec->e_cpos),
 816                                   ocfs2_rec_clusters(el, rec));
 817                 goto out;
 818         }
 819
 820         if (ret_phys_blkno)
 821                 *ret_phys_blkno = le64_to_cpu(rec->e_blkno);
 822         if (ret_cpos)
 823                 *ret_cpos = le32_to_cpu(rec->e_cpos);
 824         if (ret_clen)
 825                 *ret_clen = le16_to_cpu(rec->e_leaf_clusters);
 826
 827 out:
 828         brelse(eb_bh);
 829         return ret;
 830 }
 831
 832 /*
 833  * Returns the block index, from the start of the cluster which this
 834  * hash belongs too.
 835  */
 836 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 837                                                    u32 minor_hash)
 838 {
 839         return minor_hash & osb->osb_dx_mask;
 840 }
 841
 842 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
 843                                           struct ocfs2_dx_hinfo *hinfo)
 844 {
 845         return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
 846 }
 847
 848 static int ocfs2_dx_dir_lookup(struct inode *inode,
 849                                struct ocfs2_extent_list *el,
 850                                struct ocfs2_dx_hinfo *hinfo,
 851                                u32 *ret_cpos,
 852                                u64 *ret_phys_blkno)
 853 {
 854         int ret = 0;
 855         unsigned int cend, clen;
 856         u32 cpos;
 857         u64 blkno;
 858         u32 name_hash = hinfo->major_hash;
 859
 860         ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
 861                                       &clen);
 862         if (ret) {
 863                 mlog_errno(ret);
 864                 goto out;
 865         }
 866
 867         cend = cpos + clen;
 868         if (name_hash >= cend) {
 869                 /* We want the last cluster */
 870                 blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
 871                 cpos += clen - 1;
 872         } else {
 873                 blkno += ocfs2_clusters_to_blocks(inode->i_sb,
 874                                                   name_hash - cpos);
 875                 cpos = name_hash;
 876         }
 877
 878         /*
 879          * We now have the cluster which should hold our entry. To
 880          * find the exact block from the start of the cluster to
 881          * search, we take the lower bits of the hash.
 882          */
 883         blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
 884
 885         if (ret_phys_blkno)
 886                 *ret_phys_blkno = blkno;
 887         if (ret_cpos)
 888                 *ret_cpos = cpos;
 889
 890 out:
 891
 892         return ret;
 893 }
 894
 895 static int ocfs2_dx_dir_search(const char *name, int namelen,
 896                                struct inode *dir,
 897                                struct ocfs2_dx_root_block *dx_root,
 898                                struct ocfs2_dir_lookup_result *res)
 899 {
 900         int ret, i, found;
 901         u64 phys;
 902         struct buffer_head *dx_leaf_bh = NULL;
 903         struct ocfs2_dx_leaf *dx_leaf;
 904         struct ocfs2_dx_entry *dx_entry = NULL;
 905         struct buffer_head *dir_ent_bh = NULL;
 906         struct ocfs2_dir_entry *dir_ent = NULL;
 907         struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
 908         struct ocfs2_extent_list *dr_el;
 909         struct ocfs2_dx_entry_list *entry_list;
 910
 911         ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
 912
 913         if (ocfs2_dx_root_inline(dx_root)) {
 914                 entry_list = &dx_root->dr_entries;
 915                 goto search;
 916         }
 917
 918         dr_el = &dx_root->dr_list;
 919
 920         ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
 921         if (ret) {
 922                 mlog_errno(ret);
 923                 goto out;
 924         }
 925
 926         trace_ocfs2_dx_dir_search((unsigned long long)OCFS2_I(dir)->ip_blkno,
 927                                   namelen, name, hinfo->major_hash,
 928                                   hinfo->minor_hash, (unsigned long long)phys);
 929
 930         ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
 931         if (ret) {
 932                 mlog_errno(ret);
 933                 goto out;
 934         }
 935
 936         dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
 937
 938         trace_ocfs2_dx_dir_search_leaf_info(
 939                         le16_to_cpu(dx_leaf->dl_list.de_num_used),
 940                         le16_to_cpu(dx_leaf->dl_list.de_count));
 941
 942         entry_list = &dx_leaf->dl_list;
 943
 944 search:
 945         /*
 946          * Empty leaf is legal, so no need to check for that.
 947          */
 948         found = 0;
 949         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
 950                 dx_entry = &entry_list->de_entries[i];
 951
 952                 if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
 953                     || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
 954                         continue;
 955
 956                 /*
 957                  * Search unindexed leaf block now. We're not
 958                  * guaranteed to find anything.
 959                  */
 960                 ret = ocfs2_read_dir_block_direct(dir,
 961                                           le64_to_cpu(dx_entry->dx_dirent_blk),
 962                                           &dir_ent_bh);
 963                 if (ret) {
 964                         mlog_errno(ret);
 965                         goto out;
 966                 }
 967
 968                 /*
 969                  * XXX: We should check the unindexed block here,
 970                  * before using it.
 971                  */
 972
 973                 found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
 974                                               0, dir_ent_bh->b_data,
 975                                               dir->i_sb->s_blocksize, &dir_ent);
 976                 if (found == 1)
 977                         break;
 978
 979                 if (found == -1) {
 980                         /* This means we found a bad directory entry. */
 981                         ret = -EIO;
 982                         mlog_errno(ret);
 983                         goto out;
 984                 }
 985
 986                 brelse(dir_ent_bh);
 987                 dir_ent_bh = NULL;
 988         }
 989
 990         if (found <= 0) {
 991                 ret = -ENOENT;
 992                 goto out;
 993         }
 994
 995         res->dl_leaf_bh = dir_ent_bh;
 996         res->dl_entry = dir_ent;
 997         res->dl_dx_leaf_bh = dx_leaf_bh;
 998         res->dl_dx_entry = dx_entry;
 999
1000         ret = 0;
1001 out:
1002         if (ret) {
1003                 brelse(dx_leaf_bh);
1004                 brelse(dir_ent_bh);
1005         }
1006         return ret;
1007 }
1008
1009 static int ocfs2_find_entry_dx(const char *name, int namelen,
1010                                struct inode *dir,
1011                                struct ocfs2_dir_lookup_result *lookup)
1012 {
1013         int ret;
1014         struct buffer_head *di_bh = NULL;
1015         struct ocfs2_dinode *di;
1016         struct buffer_head *dx_root_bh = NULL;
1017         struct ocfs2_dx_root_block *dx_root;
1018
1019         ret = ocfs2_read_inode_block(dir, &di_bh);
1020         if (ret) {
1021                 mlog_errno(ret);
1022                 goto out;
1023         }
1024
1025         di = (struct ocfs2_dinode *)di_bh->b_data;
1026
1027         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1028         if (ret) {
1029                 mlog_errno(ret);
1030                 goto out;
1031         }
1032         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1033
1034         ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1035         if (ret) {
1036                 if (ret != -ENOENT)
1037                         mlog_errno(ret);
1038                 goto out;
1039         }
1040
1041         lookup->dl_dx_root_bh = dx_root_bh;
1042         dx_root_bh = NULL;
1043 out:
1044         brelse(di_bh);
1045         brelse(dx_root_bh);
1046         return ret;
1047 }
1048
1049 /*
1050  * Try to find an entry of the provided name within 'dir'.
1051  *
1052  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1053  * returned and the struct 'res' will contain information useful to
1054  * other directory manipulation functions.
1055  *
1056  * Caller can NOT assume anything about the contents of the
1057  * buffer_heads - they are passed back only so that it can be passed
1058  * into any one of the manipulation functions (add entry, delete
1059  * entry, etc). As an example, bh in the extent directory case is a
1060  * data block, in the inline-data case it actually points to an inode,
1061  * in the indexed directory case, multiple buffers are involved.
1062  */
1063 int ocfs2_find_entry(const char *name, int namelen,
1064                      struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1065 {
1066         struct buffer_head *bh;
1067         struct ocfs2_dir_entry *res_dir = NULL;
1068
1069         if (ocfs2_dir_indexed(dir))
1070                 return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1071
1072         /*
1073          * The unindexed dir code only uses part of the lookup
1074          * structure, so there's no reason to push it down further
1075          * than this.
1076          */
1077         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1078                 bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1079         else
1080                 bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1081
1082         if (bh == NULL)
1083                 return -ENOENT;
1084
1085         lookup->dl_leaf_bh = bh;
1086         lookup->dl_entry = res_dir;
1087         return 0;
1088 }
1089
1090 /*
1091  * Update inode number and type of a previously found directory entry.
1092  */
1093 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1094                        struct ocfs2_dir_lookup_result *res,
1095                        struct inode *new_entry_inode)
1096 {
1097         int ret;
1098         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1099         struct ocfs2_dir_entry *de = res->dl_entry;
1100         struct buffer_head *de_bh = res->dl_leaf_bh;
1101
1102         /*
1103          * The same code works fine for both inline-data and extent
1104          * based directories, so no need to split this up.  The only
1105          * difference is the journal_access function.
1106          */
1107
1108         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1109                 access = ocfs2_journal_access_di;
1110
1111         ret = access(handle, INODE_CACHE(dir), de_bh,
1112                      OCFS2_JOURNAL_ACCESS_WRITE);
1113         if (ret) {
1114                 mlog_errno(ret);
1115                 goto out;
1116         }
1117
1118         de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1119         ocfs2_set_de_type(de, new_entry_inode->i_mode);
1120
1121         ocfs2_journal_dirty(handle, de_bh);
1122
1123 out:
1124         return ret;
1125 }
1126
1127 /*
1128  * __ocfs2_delete_entry deletes a directory entry by merging it with the
1129  * previous entry
1130  */
1131 static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1132                                 struct ocfs2_dir_entry *de_del,
1133                                 struct buffer_head *bh, char *first_de,
1134                                 unsigned int bytes)
1135 {
1136         struct ocfs2_dir_entry *de, *pde;
1137         int i, status = -ENOENT;
1138         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1139
1140         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1141                 access = ocfs2_journal_access_di;
1142
1143         i = 0;
1144         pde = NULL;
1145         de = (struct ocfs2_dir_entry *) first_de;
1146         while (i < bytes) {
1147                 if (!ocfs2_check_dir_entry(dir, de, bh, first_de, bytes, i)) {
1148                         status = -EIO;
1149                         mlog_errno(status);
1150                         goto bail;
1151                 }
1152                 if (de == de_del)  {
1153                         status = access(handle, INODE_CACHE(dir), bh,
1154                                         OCFS2_JOURNAL_ACCESS_WRITE);
1155                         if (status < 0) {
1156                                 status = -EIO;
1157                                 mlog_errno(status);
1158                                 goto bail;
1159                         }
1160                         if (pde)
1161                                 le16_add_cpu(&pde->rec_len,
1162                                                 le16_to_cpu(de->rec_len));
1163                         de->inode = 0;
1164                         inode_inc_iversion(dir);
1165                         ocfs2_journal_dirty(handle, bh);
1166                         goto bail;
1167                 }
1168                 i += le16_to_cpu(de->rec_len);
1169                 pde = de;
1170                 de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1171         }
1172 bail:
1173         return status;
1174 }
1175
1176 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1177 {
1178         unsigned int hole;
1179
1180         if (le64_to_cpu(de->inode) == 0)
1181                 hole = le16_to_cpu(de->rec_len);
1182         else
1183                 hole = le16_to_cpu(de->rec_len) -
1184                         OCFS2_DIR_REC_LEN(de->name_len);
1185
1186         return hole;
1187 }
1188
1189 static int ocfs2_find_max_rec_len(struct super_block *sb,
1190                                   struct buffer_head *dirblock_bh)
1191 {
1192         int size, this_hole, largest_hole = 0;
1193         char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1194         struct ocfs2_dir_entry *de;
1195
1196         trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1197         size = ocfs2_dir_trailer_blk_off(sb);
1198         limit = start + size;
1199         de_buf = start;
1200         de = (struct ocfs2_dir_entry *)de_buf;
1201         do {
1202                 if (de_buf != trailer) {
1203                         this_hole = ocfs2_figure_dirent_hole(de);
1204                         if (this_hole > largest_hole)
1205                                 largest_hole = this_hole;
1206                 }
1207
1208                 de_buf += le16_to_cpu(de->rec_len);
1209                 de = (struct ocfs2_dir_entry *)de_buf;
1210         } while (de_buf < limit);
1211
1212         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1213                 return largest_hole;
1214         return 0;
1215 }
1216
1217 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1218                                        int index)
1219 {
1220         int num_used = le16_to_cpu(entry_list->de_num_used);
1221
1222         if (num_used == 1 || index == (num_used - 1))
1223                 goto clear;
1224
1225         memmove(&entry_list->de_entries[index],
1226                 &entry_list->de_entries[index + 1],
1227                 (num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1228 clear:
1229         num_used--;
1230         memset(&entry_list->de_entries[num_used], 0,
1231                sizeof(struct ocfs2_dx_entry));
1232         entry_list->de_num_used = cpu_to_le16(num_used);
1233 }
1234
1235 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1236                                  struct ocfs2_dir_lookup_result *lookup)
1237 {
1238         int ret, index, max_rec_len, add_to_free_list = 0;
1239         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1240         struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1241         struct ocfs2_dx_leaf *dx_leaf;
1242         struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1243         struct ocfs2_dir_block_trailer *trailer;
1244         struct ocfs2_dx_root_block *dx_root;
1245         struct ocfs2_dx_entry_list *entry_list;
1246
1247         /*
1248          * This function gets a bit messy because we might have to
1249          * modify the root block, regardless of whether the indexed
1250          * entries are stored inline.
1251          */
1252
1253         /*
1254          * *Only* set 'entry_list' here, based on where we're looking
1255          * for the indexed entries. Later, we might still want to
1256          * journal both blocks, based on free list state.
1257          */
1258         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1259         if (ocfs2_dx_root_inline(dx_root)) {
1260                 entry_list = &dx_root->dr_entries;
1261         } else {
1262                 dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1263                 entry_list = &dx_leaf->dl_list;
1264         }
1265
1266         /* Neither of these are a disk corruption - that should have
1267          * been caught by lookup, before we got here. */
1268         BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1269         BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1270
1271         index = (char *)dx_entry - (char *)entry_list->de_entries;
1272         index /= sizeof(*dx_entry);
1273
1274         if (index >= le16_to_cpu(entry_list->de_num_used)) {
1275                 mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1276                      (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1277                      entry_list, dx_entry);
1278                 return -EIO;
1279         }
1280
1281         /*
1282          * We know that removal of this dirent will leave enough room
1283          * for a new one, so add this block to the free list if it
1284          * isn't already there.
1285          */
1286         trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1287         if (trailer->db_free_rec_len == 0)
1288                 add_to_free_list = 1;
1289
1290         /*
1291          * Add the block holding our index into the journal before
1292          * removing the unindexed entry. If we get an error return
1293          * from __ocfs2_delete_entry(), then it hasn't removed the
1294          * entry yet. Likewise, successful return means we *must*
1295          * remove the indexed entry.
1296          *
1297          * We're also careful to journal the root tree block here as
1298          * the entry count needs to be updated. Also, we might be
1299          * adding to the start of the free list.
1300          */
1301         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1302                                       OCFS2_JOURNAL_ACCESS_WRITE);
1303         if (ret) {
1304                 mlog_errno(ret);
1305                 goto out;
1306         }
1307
1308         if (!ocfs2_dx_root_inline(dx_root)) {
1309                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1310                                               lookup->dl_dx_leaf_bh,
1311                                               OCFS2_JOURNAL_ACCESS_WRITE);
1312                 if (ret) {
1313                         mlog_errno(ret);
1314                         goto out;
1315                 }
1316         }
1317
1318         trace_ocfs2_delete_entry_dx((unsigned long long)OCFS2_I(dir)->ip_blkno,
1319                                     index);
1320
1321         ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1322                                    leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1323         if (ret) {
1324                 mlog_errno(ret);
1325                 goto out;
1326         }
1327
1328         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1329         trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1330         if (add_to_free_list) {
1331                 trailer->db_free_next = dx_root->dr_free_blk;
1332                 dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1333                 ocfs2_journal_dirty(handle, dx_root_bh);
1334         }
1335
1336         /* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1337         ocfs2_journal_dirty(handle, leaf_bh);
1338
1339         le32_add_cpu(&dx_root->dr_num_entries, -1);
1340         ocfs2_journal_dirty(handle, dx_root_bh);
1341
1342         ocfs2_dx_list_remove_entry(entry_list, index);
1343
1344         if (!ocfs2_dx_root_inline(dx_root))
1345                 ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1346
1347 out:
1348         return ret;
1349 }
1350
1351 static inline int ocfs2_delete_entry_id(handle_t *handle,
1352                                         struct inode *dir,
1353                                         struct ocfs2_dir_entry *de_del,
1354                                         struct buffer_head *bh)
1355 {
1356         int ret;
1357         struct buffer_head *di_bh = NULL;
1358         struct ocfs2_dinode *di;
1359         struct ocfs2_inline_data *data;
1360
1361         ret = ocfs2_read_inode_block(dir, &di_bh);
1362         if (ret) {
1363                 mlog_errno(ret);
1364                 goto out;
1365         }
1366
1367         di = (struct ocfs2_dinode *)di_bh->b_data;
1368         data = &di->id2.i_data;
1369
1370         ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1371                                    i_size_read(dir));
1372
1373         brelse(di_bh);
1374 out:
1375         return ret;
1376 }
1377
1378 static inline int ocfs2_delete_entry_el(handle_t *handle,
1379                                         struct inode *dir,
1380                                         struct ocfs2_dir_entry *de_del,
1381                                         struct buffer_head *bh)
1382 {
1383         return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1384                                     bh->b_size);
1385 }
1386
1387 /*
1388  * Delete a directory entry. Hide the details of directory
1389  * implementation from the caller.
1390  */
1391 int ocfs2_delete_entry(handle_t *handle,
1392                        struct inode *dir,
1393                        struct ocfs2_dir_lookup_result *res)
1394 {
1395         if (ocfs2_dir_indexed(dir))
1396                 return ocfs2_delete_entry_dx(handle, dir, res);
1397
1398         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1399                 return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1400                                              res->dl_leaf_bh);
1401
1402         return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1403                                      res->dl_leaf_bh);
1404 }
1405
1406 /*
1407  * Check whether 'de' has enough room to hold an entry of
1408  * 'new_rec_len' bytes.
1409  */
1410 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1411                                          unsigned int new_rec_len)
1412 {
1413         unsigned int de_really_used;
1414
1415         /* Check whether this is an empty record with enough space */
1416         if (le64_to_cpu(de->inode) == 0 &&
1417             le16_to_cpu(de->rec_len) >= new_rec_len)
1418                 return 1;
1419
1420         /*
1421          * Record might have free space at the end which we can
1422          * use.
1423          */
1424         de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1425         if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1426             return 1;
1427
1428         return 0;
1429 }
1430
1431 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1432                                           struct ocfs2_dx_entry *dx_new_entry)
1433 {
1434         int i;
1435
1436         i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1437         dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1438
1439         le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1440 }
1441
1442 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1443                                        struct ocfs2_dx_hinfo *hinfo,
1444                                        u64 dirent_blk)
1445 {
1446         int i;
1447         struct ocfs2_dx_entry *dx_entry;
1448
1449         i = le16_to_cpu(entry_list->de_num_used);
1450         dx_entry = &entry_list->de_entries[i];
1451
1452         memset(dx_entry, 0, sizeof(*dx_entry));
1453         dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1454         dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1455         dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1456
1457         le16_add_cpu(&entry_list->de_num_used, 1);
1458 }
1459
1460 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1461                                       struct ocfs2_dx_hinfo *hinfo,
1462                                       u64 dirent_blk,
1463                                       struct buffer_head *dx_leaf_bh)
1464 {
1465         int ret;
1466         struct ocfs2_dx_leaf *dx_leaf;
1467
1468         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1469                                       OCFS2_JOURNAL_ACCESS_WRITE);
1470         if (ret) {
1471                 mlog_errno(ret);
1472                 goto out;
1473         }
1474
1475         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1476         ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1477         ocfs2_journal_dirty(handle, dx_leaf_bh);
1478
1479 out:
1480         return ret;
1481 }
1482
1483 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1484                                         struct ocfs2_dx_hinfo *hinfo,
1485                                         u64 dirent_blk,
1486                                         struct ocfs2_dx_root_block *dx_root)
1487 {
1488         ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1489 }
1490
1491 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1492                                struct ocfs2_dir_lookup_result *lookup)
1493 {
1494         int ret = 0;
1495         struct ocfs2_dx_root_block *dx_root;
1496         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1497
1498         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1499                                       OCFS2_JOURNAL_ACCESS_WRITE);
1500         if (ret) {
1501                 mlog_errno(ret);
1502                 goto out;
1503         }
1504
1505         dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1506         if (ocfs2_dx_root_inline(dx_root)) {
1507                 ocfs2_dx_inline_root_insert(dir, handle,
1508                                             &lookup->dl_hinfo,
1509                                             lookup->dl_leaf_bh->b_blocknr,
1510                                             dx_root);
1511         } else {
1512                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1513                                                  lookup->dl_leaf_bh->b_blocknr,
1514                                                  lookup->dl_dx_leaf_bh);
1515                 if (ret)
1516                         goto out;
1517         }
1518
1519         le32_add_cpu(&dx_root->dr_num_entries, 1);
1520         ocfs2_journal_dirty(handle, dx_root_bh);
1521
1522 out:
1523         return ret;
1524 }
1525
1526 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1527                                        handle_t *handle,
1528                                        struct ocfs2_dir_lookup_result *lookup)
1529 {
1530         struct ocfs2_dir_block_trailer *trailer, *prev;
1531         struct ocfs2_dx_root_block *dx_root;
1532         struct buffer_head *bh;
1533
1534         trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1535
1536         if (ocfs2_free_list_at_root(lookup)) {
1537                 bh = lookup->dl_dx_root_bh;
1538                 dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1539                 dx_root->dr_free_blk = trailer->db_free_next;
1540         } else {
1541                 bh = lookup->dl_prev_leaf_bh;
1542                 prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1543                 prev->db_free_next = trailer->db_free_next;
1544         }
1545
1546         trailer->db_free_rec_len = cpu_to_le16(0);
1547         trailer->db_free_next = cpu_to_le64(0);
1548
1549         ocfs2_journal_dirty(handle, bh);
1550         ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1551 }
1552
1553 /*
1554  * This expects that a journal write has been reserved on
1555  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1556  */
1557 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1558                                    struct ocfs2_dir_lookup_result *lookup)
1559 {
1560         int max_rec_len;
1561         struct ocfs2_dir_block_trailer *trailer;
1562
1563         /* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1564         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1565         if (max_rec_len) {
1566                 /*
1567                  * There's still room in this block, so no need to remove it
1568                  * from the free list. In this case, we just want to update
1569                  * the rec len accounting.
1570                  */
1571                 trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1572                 trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1573                 ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1574         } else {
1575                 ocfs2_remove_block_from_free_list(dir, handle, lookup);
1576         }
1577 }
1578
1579 /* we don't always have a dentry for what we want to add, so people
1580  * like orphan dir can call this instead.
1581  *
1582  * The lookup context must have been filled from
1583  * ocfs2_prepare_dir_for_insert.
1584  */
1585 int __ocfs2_add_entry(handle_t *handle,
1586                       struct inode *dir,
1587                       const char *name, int namelen,
1588                       struct inode *inode, u64 blkno,
1589                       struct buffer_head *parent_fe_bh,
1590                       struct ocfs2_dir_lookup_result *lookup)
1591 {
1592         unsigned long offset;
1593         unsigned short rec_len;
1594         struct ocfs2_dir_entry *de, *de1;
1595         struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1596         struct super_block *sb = dir->i_sb;
1597         int retval;
1598         unsigned int size = sb->s_blocksize;
1599         struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1600         char *data_start = insert_bh->b_data;
1601
1602         if (ocfs2_dir_indexed(dir)) {
1603                 struct buffer_head *bh;
1604
1605                 /*
1606                  * An indexed dir may require that we update the free space
1607                  * list. Reserve a write to the previous node in the list so
1608                  * that we don't fail later.
1609                  *
1610                  * XXX: This can be either a dx_root_block, or an unindexed
1611                  * directory tree leaf block.
1612                  */
1613                 if (ocfs2_free_list_at_root(lookup)) {
1614                         bh = lookup->dl_dx_root_bh;
1615                         retval = ocfs2_journal_access_dr(handle,
1616                                                  INODE_CACHE(dir), bh,
1617                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1618                 } else {
1619                         bh = lookup->dl_prev_leaf_bh;
1620                         retval = ocfs2_journal_access_db(handle,
1621                                                  INODE_CACHE(dir), bh,
1622                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1623                 }
1624                 if (retval) {
1625                         mlog_errno(retval);
1626                         return retval;
1627                 }
1628         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1629                 data_start = di->id2.i_data.id_data;
1630                 size = i_size_read(dir);
1631
1632                 BUG_ON(insert_bh != parent_fe_bh);
1633         }
1634
1635         rec_len = OCFS2_DIR_REC_LEN(namelen);
1636         offset = 0;
1637         de = (struct ocfs2_dir_entry *) data_start;
1638         while (1) {
1639                 BUG_ON((char *)de >= (size + data_start));
1640
1641                 /* These checks should've already been passed by the
1642                  * prepare function, but I guess we can leave them
1643                  * here anyway. */
1644                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, data_start,
1645                                            size, offset)) {
1646                         retval = -ENOENT;
1647                         goto bail;
1648                 }
1649                 if (ocfs2_match(namelen, name, de)) {
1650                         retval = -EEXIST;
1651                         goto bail;
1652                 }
1653
1654                 /* We're guaranteed that we should have space, so we
1655                  * can't possibly have hit the trailer...right? */
1656                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1657                                 "Hit dir trailer trying to insert %.*s "
1658                                 "(namelen %d) into directory %llu.  "
1659                                 "offset is %lu, trailer offset is %d\n",
1660                                 namelen, name, namelen,
1661                                 (unsigned long long)parent_fe_bh->b_blocknr,
1662                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1663
1664                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1665                         inode_set_mtime_to_ts(dir,
1666                                               inode_set_ctime_current(dir));
1667                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1668                         if (retval < 0) {
1669                                 mlog_errno(retval);
1670                                 goto bail;
1671                         }
1672
1673                         if (insert_bh == parent_fe_bh)
1674                                 retval = ocfs2_journal_access_di(handle,
1675                                                                  INODE_CACHE(dir),
1676                                                                  insert_bh,
1677                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1678                         else {
1679                                 retval = ocfs2_journal_access_db(handle,
1680                                                                  INODE_CACHE(dir),
1681                                                                  insert_bh,
1682                                               OCFS2_JOURNAL_ACCESS_WRITE);
1683
1684                                 if (!retval && ocfs2_dir_indexed(dir))
1685                                         retval = ocfs2_dx_dir_insert(dir,
1686                                                                 handle,
1687                                                                 lookup);
1688                         }
1689
1690                         if (retval) {
1691                                 mlog_errno(retval);
1692                                 goto bail;
1693                         }
1694
1695                         /* By now the buffer is marked for journaling */
1696                         offset += le16_to_cpu(de->rec_len);
1697                         if (le64_to_cpu(de->inode)) {
1698                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1699                                         OCFS2_DIR_REC_LEN(de->name_len));
1700                                 de1->rec_len =
1701                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1702                                         OCFS2_DIR_REC_LEN(de->name_len));
1703                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1704                                 de = de1;
1705                         }
1706                         de->file_type = FT_UNKNOWN;
1707                         if (blkno) {
1708                                 de->inode = cpu_to_le64(blkno);
1709                                 ocfs2_set_de_type(de, inode->i_mode);
1710                         } else
1711                                 de->inode = 0;
1712                         de->name_len = namelen;
1713                         memcpy(de->name, name, namelen);
1714
1715                         if (ocfs2_dir_indexed(dir))
1716                                 ocfs2_recalc_free_list(dir, handle, lookup);
1717
1718                         inode_inc_iversion(dir);
1719                         ocfs2_journal_dirty(handle, insert_bh);
1720                         retval = 0;
1721                         goto bail;
1722                 }
1723
1724                 offset += le16_to_cpu(de->rec_len);
1725                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1726         }
1727
1728         /* when you think about it, the assert above should prevent us
1729          * from ever getting here. */
1730         retval = -ENOSPC;
1731 bail:
1732         if (retval)
1733                 mlog_errno(retval);
1734
1735         return retval;
1736 }
1737
1738 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1739                                     u64 *f_version,
1740                                     struct dir_context *ctx)
1741 {
1742         int ret, i;
1743         unsigned long offset = ctx->pos;
1744         struct buffer_head *di_bh = NULL;
1745         struct ocfs2_dinode *di;
1746         struct ocfs2_inline_data *data;
1747         struct ocfs2_dir_entry *de;
1748
1749         ret = ocfs2_read_inode_block(inode, &di_bh);
1750         if (ret) {
1751                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1752                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1753                 goto out;
1754         }
1755
1756         di = (struct ocfs2_dinode *)di_bh->b_data;
1757         data = &di->id2.i_data;
1758
1759         while (ctx->pos < i_size_read(inode)) {
1760                 /* If the dir block has changed since the last call to
1761                  * readdir(2), then we might be pointing to an invalid
1762                  * dirent right now.  Scan from the start of the block
1763                  * to make sure. */
1764                 if (!inode_eq_iversion(inode, *f_version)) {
1765                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1766                                 de = (struct ocfs2_dir_entry *)
1767                                         (data->id_data + i);
1768                                 /* It's too expensive to do a full
1769                                  * dirent test each time round this
1770                                  * loop, but we do have to test at
1771                                  * least that it is non-zero.  A
1772                                  * failure will be detected in the
1773                                  * dirent test below. */
1774                                 if (le16_to_cpu(de->rec_len) <
1775                                     OCFS2_DIR_REC_LEN(1))
1776                                         break;
1777                                 i += le16_to_cpu(de->rec_len);
1778                         }
1779                         ctx->pos = offset = i;
1780                         *f_version = inode_query_iversion(inode);
1781                 }
1782
1783                 de = (struct ocfs2_dir_entry *) (data->id_data + ctx->pos);
1784                 if (!ocfs2_check_dir_entry(inode, de, di_bh, (char *)data->id_data,
1785                                            i_size_read(inode), ctx->pos)) {
1786                         /* On error, skip the f_pos to the end. */
1787                         ctx->pos = i_size_read(inode);
1788                         break;
1789                 }
1790                 offset += le16_to_cpu(de->rec_len);
1791                 if (le64_to_cpu(de->inode)) {
1792                         if (!dir_emit(ctx, de->name, de->name_len,
1793                                       le64_to_cpu(de->inode),
1794                                       fs_ftype_to_dtype(de->file_type)))
1795                                 goto out;
1796                 }
1797                 ctx->pos += le16_to_cpu(de->rec_len);
1798         }
1799 out:
1800         brelse(di_bh);
1801         return 0;
1802 }
1803
1804 /*
1805  * NOTE: This function can be called against unindexed directories,
1806  * and indexed ones.
1807  */
1808 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1809                                     u64 *f_version,
1810                                     struct dir_context *ctx,
1811                                     bool persist)
1812 {
1813         unsigned long offset, blk, last_ra_blk = 0;
1814         int i;
1815         struct buffer_head * bh, * tmp;
1816         struct ocfs2_dir_entry * de;
1817         struct super_block * sb = inode->i_sb;
1818         unsigned int ra_sectors = 16;
1819         int stored = 0;
1820
1821         bh = NULL;
1822
1823         offset = ctx->pos & (sb->s_blocksize - 1);
1824
1825         while (ctx->pos < i_size_read(inode)) {
1826                 blk = ctx->pos >> sb->s_blocksize_bits;
1827                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1828                         /* Skip the corrupt dirblock and keep trying */
1829                         ctx->pos += sb->s_blocksize - offset;
1830                         continue;
1831                 }
1832
1833                 /* The idea here is to begin with 8k read-ahead and to stay
1834                  * 4k ahead of our current position.
1835                  *
1836                  * TODO: Use the pagecache for this. We just need to
1837                  * make sure it's cluster-safe... */
1838                 if (!last_ra_blk
1839                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1840                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1841                              i > 0; i--) {
1842                                 tmp = NULL;
1843                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1844                                                           OCFS2_BH_READAHEAD))
1845                                         brelse(tmp);
1846                         }
1847                         last_ra_blk = blk;
1848                         ra_sectors = 8;
1849                 }
1850
1851                 /* If the dir block has changed since the last call to
1852                  * readdir(2), then we might be pointing to an invalid
1853                  * dirent right now.  Scan from the start of the block
1854                  * to make sure. */
1855                 if (!inode_eq_iversion(inode, *f_version)) {
1856                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1857                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1858                                 /* It's too expensive to do a full
1859                                  * dirent test each time round this
1860                                  * loop, but we do have to test at
1861                                  * least that it is non-zero.  A
1862                                  * failure will be detected in the
1863                                  * dirent test below. */
1864                                 if (le16_to_cpu(de->rec_len) <
1865                                     OCFS2_DIR_REC_LEN(1))
1866                                         break;
1867                                 i += le16_to_cpu(de->rec_len);
1868                         }
1869                         offset = i;
1870                         ctx->pos = (ctx->pos & ~(sb->s_blocksize - 1))
1871                                 | offset;
1872                         *f_version = inode_query_iversion(inode);
1873                 }
1874
1875                 while (ctx->pos < i_size_read(inode)
1876                        && offset < sb->s_blocksize) {
1877                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1878                         if (!ocfs2_check_dir_entry(inode, de, bh, bh->b_data,
1879                                                    sb->s_blocksize, offset)) {
1880                                 /* On error, skip the f_pos to the
1881                                    next block. */
1882                                 ctx->pos = (ctx->pos | (sb->s_blocksize - 1)) + 1;
1883                                 break;
1884                         }
1885                         if (le64_to_cpu(de->inode)) {
1886                                 if (!dir_emit(ctx, de->name,
1887                                                 de->name_len,
1888                                                 le64_to_cpu(de->inode),
1889                                         fs_ftype_to_dtype(de->file_type))) {
1890                                         brelse(bh);
1891                                         return 0;
1892                                 }
1893                                 stored++;
1894                         }
1895                         offset += le16_to_cpu(de->rec_len);
1896                         ctx->pos += le16_to_cpu(de->rec_len);
1897                 }
1898                 offset = 0;
1899                 brelse(bh);
1900                 bh = NULL;
1901                 if (!persist && stored)
1902                         break;
1903         }
1904         return 0;
1905 }
1906
1907 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1908                                  struct dir_context *ctx,
1909                                  bool persist)
1910 {
1911         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1912                 return ocfs2_dir_foreach_blk_id(inode, f_version, ctx);
1913         return ocfs2_dir_foreach_blk_el(inode, f_version, ctx, persist);
1914 }
1915
1916 /*
1917  * This is intended to be called from inside other kernel functions,
1918  * so we fake some arguments.
1919  */
1920 int ocfs2_dir_foreach(struct inode *inode, struct dir_context *ctx)
1921 {
1922         u64 version = inode_query_iversion(inode);
1923         ocfs2_dir_foreach_blk(inode, &version, ctx, true);
1924         return 0;
1925 }
1926
1927 /*
1928  * ocfs2_readdir()
1929  *
1930  */
1931 int ocfs2_readdir(struct file *file, struct dir_context *ctx)
1932 {
1933         int error = 0;
1934         struct inode *inode = file_inode(file);
1935         int lock_level = 0;
1936
1937         trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
1938
1939         error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level, 1);
1940         if (lock_level && error >= 0) {
1941                 /* We release EX lock which used to update atime
1942                  * and get PR lock again to reduce contention
1943                  * on commonly accessed directories. */
1944                 ocfs2_inode_unlock(inode, 1);
1945                 lock_level = 0;
1946                 error = ocfs2_inode_lock(inode, NULL, 0);
1947         }
1948         if (error < 0) {
1949                 if (error != -ENOENT)
1950                         mlog_errno(error);
1951                 /* we haven't got any yet, so propagate the error. */
1952                 goto bail_nolock;
1953         }
1954
1955         error = ocfs2_dir_foreach_blk(inode, &file->f_version, ctx, false);
1956
1957         ocfs2_inode_unlock(inode, lock_level);
1958         if (error)
1959                 mlog_errno(error);
1960
1961 bail_nolock:
1962
1963         return error;
1964 }
1965
1966 /*
1967  * NOTE: this should always be called with parent dir i_rwsem taken.
1968  */
1969 int ocfs2_find_files_on_disk(const char *name,
1970                              int namelen,
1971                              u64 *blkno,
1972                              struct inode *inode,
1973                              struct ocfs2_dir_lookup_result *lookup)
1974 {
1975         int status = -ENOENT;
1976
1977         trace_ocfs2_find_files_on_disk(namelen, name, blkno,
1978                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1979
1980         status = ocfs2_find_entry(name, namelen, inode, lookup);
1981         if (status)
1982                 goto leave;
1983
1984         *blkno = le64_to_cpu(lookup->dl_entry->inode);
1985
1986         status = 0;
1987 leave:
1988
1989         return status;
1990 }
1991
1992 /*
1993  * Convenience function for callers which just want the block number
1994  * mapped to a name and don't require the full dirent info, etc.
1995  */
1996 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
1997                                int namelen, u64 *blkno)
1998 {
1999         int ret;
2000         struct ocfs2_dir_lookup_result lookup = { NULL, };
2001
2002         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2003         ocfs2_free_dir_lookup_result(&lookup);
2004
2005         return ret;
2006 }
2007
2008 /* Check for a name within a directory.
2009  *
2010  * Return 0 if the name does not exist
2011  * Return -EEXIST if the directory contains the name
2012  *
2013  * Callers should have i_rwsem + a cluster lock on dir
2014  */
2015 int ocfs2_check_dir_for_entry(struct inode *dir,
2016                               const char *name,
2017                               int namelen)
2018 {
2019         int ret = 0;
2020         struct ocfs2_dir_lookup_result lookup = { NULL, };
2021
2022         trace_ocfs2_check_dir_for_entry(
2023                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2024
2025         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0) {
2026                 ret = -EEXIST;
2027                 mlog_errno(ret);
2028         }
2029
2030         ocfs2_free_dir_lookup_result(&lookup);
2031
2032         return ret;
2033 }
2034
2035 struct ocfs2_empty_dir_priv {
2036         struct dir_context ctx;
2037         unsigned seen_dot;
2038         unsigned seen_dot_dot;
2039         unsigned seen_other;
2040         unsigned dx_dir;
2041 };
2042 static bool ocfs2_empty_dir_filldir(struct dir_context *ctx, const char *name,
2043                                    int name_len, loff_t pos, u64 ino,
2044                                    unsigned type)
2045 {
2046         struct ocfs2_empty_dir_priv *p =
2047                 container_of(ctx, struct ocfs2_empty_dir_priv, ctx);
2048
2049         /*
2050          * Check the positions of "." and ".." records to be sure
2051          * they're in the correct place.
2052          *
2053          * Indexed directories don't need to proceed past the first
2054          * two entries, so we end the scan after seeing '..'. Despite
2055          * that, we allow the scan to proceed In the event that we
2056          * have a corrupted indexed directory (no dot or dot dot
2057          * entries). This allows us to double check for existing
2058          * entries which might not have been found in the index.
2059          */
2060         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2061                 p->seen_dot = 1;
2062                 return true;
2063         }
2064
2065         if (name_len == 2 && !strncmp("..", name, 2) &&
2066             pos == OCFS2_DIR_REC_LEN(1)) {
2067                 p->seen_dot_dot = 1;
2068
2069                 if (p->dx_dir && p->seen_dot)
2070                         return false;
2071
2072                 return true;
2073         }
2074
2075         p->seen_other = 1;
2076         return false;
2077 }
2078
2079 static int ocfs2_empty_dir_dx(struct inode *inode,
2080                               struct ocfs2_empty_dir_priv *priv)
2081 {
2082         int ret;
2083         struct buffer_head *di_bh = NULL;
2084         struct buffer_head *dx_root_bh = NULL;
2085         struct ocfs2_dinode *di;
2086         struct ocfs2_dx_root_block *dx_root;
2087
2088         priv->dx_dir = 1;
2089
2090         ret = ocfs2_read_inode_block(inode, &di_bh);
2091         if (ret) {
2092                 mlog_errno(ret);
2093                 goto out;
2094         }
2095         di = (struct ocfs2_dinode *)di_bh->b_data;
2096
2097         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2098         if (ret) {
2099                 mlog_errno(ret);
2100                 goto out;
2101         }
2102         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2103
2104         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2105                 priv->seen_other = 1;
2106
2107 out:
2108         brelse(di_bh);
2109         brelse(dx_root_bh);
2110         return ret;
2111 }
2112
2113 /*
2114  * routine to check that the specified directory is empty (for rmdir)
2115  *
2116  * Returns 1 if dir is empty, zero otherwise.
2117  *
2118  * XXX: This is a performance problem for unindexed directories.
2119  */
2120 int ocfs2_empty_dir(struct inode *inode)
2121 {
2122         int ret;
2123         struct ocfs2_empty_dir_priv priv = {
2124                 .ctx.actor = ocfs2_empty_dir_filldir,
2125         };
2126
2127         if (ocfs2_dir_indexed(inode)) {
2128                 ret = ocfs2_empty_dir_dx(inode, &priv);
2129                 if (ret)
2130                         mlog_errno(ret);
2131                 /*
2132                  * We still run ocfs2_dir_foreach to get the checks
2133                  * for "." and "..".
2134                  */
2135         }
2136
2137         ret = ocfs2_dir_foreach(inode, &priv.ctx);
2138         if (ret)
2139                 mlog_errno(ret);
2140
2141         if (!priv.seen_dot || !priv.seen_dot_dot) {
2142                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2143                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2144                 /*
2145                  * XXX: Is it really safe to allow an unlink to continue?
2146                  */
2147                 return 1;
2148         }
2149
2150         return !priv.seen_other;
2151 }
2152
2153 /*
2154  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2155  * "..", which might be used during creation of a directory with a trailing
2156  * header. It is otherwise safe to ignore the return code.
2157  */
2158 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2159                                                           struct inode *parent,
2160                                                           char *start,
2161                                                           unsigned int size)
2162 {
2163         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2164
2165         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2166         de->name_len = 1;
2167         de->rec_len =
2168                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2169         strcpy(de->name, ".");
2170         ocfs2_set_de_type(de, S_IFDIR);
2171
2172         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2173         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2174         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2175         de->name_len = 2;
2176         strcpy(de->name, "..");
2177         ocfs2_set_de_type(de, S_IFDIR);
2178
2179         return de;
2180 }
2181
2182 /*
2183  * This works together with code in ocfs2_mknod_locked() which sets
2184  * the inline-data flag and initializes the inline-data section.
2185  */
2186 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2187                                  handle_t *handle,
2188                                  struct inode *parent,
2189                                  struct inode *inode,
2190                                  struct buffer_head *di_bh)
2191 {
2192         int ret;
2193         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2194         struct ocfs2_inline_data *data = &di->id2.i_data;
2195         unsigned int size = le16_to_cpu(data->id_count);
2196
2197         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2198                                       OCFS2_JOURNAL_ACCESS_WRITE);
2199         if (ret) {
2200                 mlog_errno(ret);
2201                 goto out;
2202         }
2203
2204         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2205         ocfs2_journal_dirty(handle, di_bh);
2206
2207         i_size_write(inode, size);
2208         set_nlink(inode, 2);
2209         inode->i_blocks = ocfs2_inode_sector_count(inode);
2210
2211         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2212         if (ret < 0)
2213                 mlog_errno(ret);
2214
2215 out:
2216         return ret;
2217 }
2218
2219 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2220                                  handle_t *handle,
2221                                  struct inode *parent,
2222                                  struct inode *inode,
2223                                  struct buffer_head *fe_bh,
2224                                  struct ocfs2_alloc_context *data_ac,
2225                                  struct buffer_head **ret_new_bh)
2226 {
2227         int status;
2228         unsigned int size = osb->sb->s_blocksize;
2229         struct buffer_head *new_bh = NULL;
2230         struct ocfs2_dir_entry *de;
2231
2232         if (ocfs2_new_dir_wants_trailer(inode))
2233                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2234
2235         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2236                                      data_ac, NULL, &new_bh);
2237         if (status < 0) {
2238                 mlog_errno(status);
2239                 goto bail;
2240         }
2241
2242         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2243
2244         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2245                                          OCFS2_JOURNAL_ACCESS_CREATE);
2246         if (status < 0) {
2247                 mlog_errno(status);
2248                 goto bail;
2249         }
2250         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2251
2252         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2253         if (ocfs2_new_dir_wants_trailer(inode)) {
2254                 int size = le16_to_cpu(de->rec_len);
2255
2256                 /*
2257                  * Figure out the size of the hole left over after
2258                  * insertion of '.' and '..'. The trailer wants this
2259                  * information.
2260                  */
2261                 size -= OCFS2_DIR_REC_LEN(2);
2262                 size -= sizeof(struct ocfs2_dir_block_trailer);
2263
2264                 ocfs2_init_dir_trailer(inode, new_bh, size);
2265         }
2266
2267         ocfs2_journal_dirty(handle, new_bh);
2268
2269         i_size_write(inode, inode->i_sb->s_blocksize);
2270         set_nlink(inode, 2);
2271         inode->i_blocks = ocfs2_inode_sector_count(inode);
2272         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2273         if (status < 0) {
2274                 mlog_errno(status);
2275                 goto bail;
2276         }
2277
2278         status = 0;
2279         if (ret_new_bh) {
2280                 *ret_new_bh = new_bh;
2281                 new_bh = NULL;
2282         }
2283 bail:
2284         brelse(new_bh);
2285
2286         return status;
2287 }
2288
2289 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2290                                      handle_t *handle, struct inode *dir,
2291                                      struct buffer_head *di_bh,
2292                                      struct buffer_head *dirdata_bh,
2293                                      struct ocfs2_alloc_context *meta_ac,
2294                                      int dx_inline, u32 num_entries,
2295                                      struct buffer_head **ret_dx_root_bh)
2296 {
2297         int ret;
2298         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2299         u16 dr_suballoc_bit;
2300         u64 suballoc_loc, dr_blkno;
2301         unsigned int num_bits;
2302         struct buffer_head *dx_root_bh = NULL;
2303         struct ocfs2_dx_root_block *dx_root;
2304         struct ocfs2_dir_block_trailer *trailer =
2305                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2306
2307         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2308                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2309         if (ret) {
2310                 mlog_errno(ret);
2311                 goto out;
2312         }
2313
2314         trace_ocfs2_dx_dir_attach_index(
2315                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2316                                 (unsigned long long)dr_blkno);
2317
2318         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2319         if (dx_root_bh == NULL) {
2320                 ret = -ENOMEM;
2321                 goto out;
2322         }
2323         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2324
2325         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2326                                       OCFS2_JOURNAL_ACCESS_CREATE);
2327         if (ret < 0) {
2328                 mlog_errno(ret);
2329                 goto out;
2330         }
2331
2332         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2333         memset(dx_root, 0, osb->sb->s_blocksize);
2334         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2335         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2336         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2337         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2338         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2339         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2340         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2341         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2342         if (le16_to_cpu(trailer->db_free_rec_len))
2343                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2344         else
2345                 dx_root->dr_free_blk = cpu_to_le64(0);
2346
2347         if (dx_inline) {
2348                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2349                 dx_root->dr_entries.de_count =
2350                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2351         } else {
2352                 dx_root->dr_list.l_count =
2353                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2354         }
2355         ocfs2_journal_dirty(handle, dx_root_bh);
2356
2357         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2358                                       OCFS2_JOURNAL_ACCESS_CREATE);
2359         if (ret) {
2360                 mlog_errno(ret);
2361                 goto out;
2362         }
2363
2364         di->i_dx_root = cpu_to_le64(dr_blkno);
2365
2366         spin_lock(&OCFS2_I(dir)->ip_lock);
2367         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2368         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2369         spin_unlock(&OCFS2_I(dir)->ip_lock);
2370
2371         ocfs2_journal_dirty(handle, di_bh);
2372
2373         *ret_dx_root_bh = dx_root_bh;
2374         dx_root_bh = NULL;
2375
2376 out:
2377         brelse(dx_root_bh);
2378         return ret;
2379 }
2380
2381 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2382                                        handle_t *handle, struct inode *dir,
2383                                        struct buffer_head **dx_leaves,
2384                                        int num_dx_leaves, u64 start_blk)
2385 {
2386         int ret, i;
2387         struct ocfs2_dx_leaf *dx_leaf;
2388         struct buffer_head *bh;
2389
2390         for (i = 0; i < num_dx_leaves; i++) {
2391                 bh = sb_getblk(osb->sb, start_blk + i);
2392                 if (bh == NULL) {
2393                         ret = -ENOMEM;
2394                         goto out;
2395                 }
2396                 dx_leaves[i] = bh;
2397
2398                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2399
2400                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2401                                               OCFS2_JOURNAL_ACCESS_CREATE);
2402                 if (ret < 0) {
2403                         mlog_errno(ret);
2404                         goto out;
2405                 }
2406
2407                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2408
2409                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2410                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2411                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2412                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2413                 dx_leaf->dl_list.de_count =
2414                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2415
2416                 trace_ocfs2_dx_dir_format_cluster(
2417                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2418                                 (unsigned long long)bh->b_blocknr,
2419                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2420
2421                 ocfs2_journal_dirty(handle, bh);
2422         }
2423
2424         ret = 0;
2425 out:
2426         return ret;
2427 }
2428
2429 /*
2430  * Allocates and formats a new cluster for use in an indexed dir
2431  * leaf. This version will not do the extent insert, so that it can be
2432  * used by operations which need careful ordering.
2433  */
2434 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2435                                       u32 cpos, handle_t *handle,
2436                                       struct ocfs2_alloc_context *data_ac,
2437                                       struct buffer_head **dx_leaves,
2438                                       int num_dx_leaves, u64 *ret_phys_blkno)
2439 {
2440         int ret;
2441         u32 phys, num;
2442         u64 phys_blkno;
2443         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2444
2445         /*
2446          * XXX: For create, this should claim cluster for the index
2447          * *before* the unindexed insert so that we have a better
2448          * chance of contiguousness as the directory grows in number
2449          * of entries.
2450          */
2451         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2452         if (ret) {
2453                 mlog_errno(ret);
2454                 goto out;
2455         }
2456
2457         /*
2458          * Format the new cluster first. That way, we're inserting
2459          * valid data.
2460          */
2461         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2462         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2463                                           num_dx_leaves, phys_blkno);
2464         if (ret) {
2465                 mlog_errno(ret);
2466                 goto out;
2467         }
2468
2469         *ret_phys_blkno = phys_blkno;
2470 out:
2471         return ret;
2472 }
2473
2474 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2475                                     struct ocfs2_extent_tree *et,
2476                                     u32 cpos, handle_t *handle,
2477                                     struct ocfs2_alloc_context *data_ac,
2478                                     struct ocfs2_alloc_context *meta_ac,
2479                                     struct buffer_head **dx_leaves,
2480                                     int num_dx_leaves)
2481 {
2482         int ret;
2483         u64 phys_blkno;
2484
2485         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2486                                          num_dx_leaves, &phys_blkno);
2487         if (ret) {
2488                 mlog_errno(ret);
2489                 goto out;
2490         }
2491
2492         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2493                                   meta_ac);
2494         if (ret)
2495                 mlog_errno(ret);
2496 out:
2497         return ret;
2498 }
2499
2500 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2501                                                         int *ret_num_leaves)
2502 {
2503         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2504         struct buffer_head **dx_leaves;
2505
2506         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2507                             GFP_NOFS);
2508         if (dx_leaves && ret_num_leaves)
2509                 *ret_num_leaves = num_dx_leaves;
2510
2511         return dx_leaves;
2512 }
2513
2514 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2515                                  handle_t *handle,
2516                                  struct inode *parent,
2517                                  struct inode *inode,
2518                                  struct buffer_head *di_bh,
2519                                  struct ocfs2_alloc_context *data_ac,
2520                                  struct ocfs2_alloc_context *meta_ac)
2521 {
2522         int ret;
2523         struct buffer_head *leaf_bh = NULL;
2524         struct buffer_head *dx_root_bh = NULL;
2525         struct ocfs2_dx_hinfo hinfo;
2526         struct ocfs2_dx_root_block *dx_root;
2527         struct ocfs2_dx_entry_list *entry_list;
2528
2529         /*
2530          * Our strategy is to create the directory as though it were
2531          * unindexed, then add the index block. This works with very
2532          * little complication since the state of a new directory is a
2533          * very well known quantity.
2534          *
2535          * Essentially, we have two dirents ("." and ".."), in the 1st
2536          * block which need indexing. These are easily inserted into
2537          * the index block.
2538          */
2539
2540         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2541                                     data_ac, &leaf_bh);
2542         if (ret) {
2543                 mlog_errno(ret);
2544                 goto out;
2545         }
2546
2547         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2548                                         meta_ac, 1, 2, &dx_root_bh);
2549         if (ret) {
2550                 mlog_errno(ret);
2551                 goto out;
2552         }
2553         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2554         entry_list = &dx_root->dr_entries;
2555
2556         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2557         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2558         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2559
2560         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2561         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2562
2563 out:
2564         brelse(dx_root_bh);
2565         brelse(leaf_bh);
2566         return ret;
2567 }
2568
2569 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2570                        handle_t *handle,
2571                        struct inode *parent,
2572                        struct inode *inode,
2573                        struct buffer_head *fe_bh,
2574                        struct ocfs2_alloc_context *data_ac,
2575                        struct ocfs2_alloc_context *meta_ac)
2576
2577 {
2578         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2579
2580         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2581                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2582
2583         if (ocfs2_supports_indexed_dirs(osb))
2584                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2585                                              data_ac, meta_ac);
2586
2587         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2588                                      data_ac, NULL);
2589 }
2590
2591 static int ocfs2_dx_dir_index_block(struct inode *dir,
2592                                     handle_t *handle,
2593                                     struct buffer_head **dx_leaves,
2594                                     int num_dx_leaves,
2595                                     u32 *num_dx_entries,
2596                                     struct buffer_head *dirent_bh)
2597 {
2598         int ret = 0, namelen, i;
2599         char *de_buf, *limit;
2600         struct ocfs2_dir_entry *de;
2601         struct buffer_head *dx_leaf_bh;
2602         struct ocfs2_dx_hinfo hinfo;
2603         u64 dirent_blk = dirent_bh->b_blocknr;
2604
2605         de_buf = dirent_bh->b_data;
2606         limit = de_buf + dir->i_sb->s_blocksize;
2607
2608         while (de_buf < limit) {
2609                 de = (struct ocfs2_dir_entry *)de_buf;
2610
2611                 namelen = de->name_len;
2612                 if (!namelen || !de->inode)
2613                         goto inc;
2614
2615                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2616
2617                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2618                 dx_leaf_bh = dx_leaves[i];
2619
2620                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2621                                                  dirent_blk, dx_leaf_bh);
2622                 if (ret) {
2623                         mlog_errno(ret);
2624                         goto out;
2625                 }
2626
2627                 *num_dx_entries = *num_dx_entries + 1;
2628
2629 inc:
2630                 de_buf += le16_to_cpu(de->rec_len);
2631         }
2632
2633 out:
2634         return ret;
2635 }
2636
2637 /*
2638  * XXX: This expects dx_root_bh to already be part of the transaction.
2639  */
2640 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2641                                          struct buffer_head *dx_root_bh,
2642                                          struct buffer_head *dirent_bh)
2643 {
2644         char *de_buf, *limit;
2645         struct ocfs2_dx_root_block *dx_root;
2646         struct ocfs2_dir_entry *de;
2647         struct ocfs2_dx_hinfo hinfo;
2648         u64 dirent_blk = dirent_bh->b_blocknr;
2649
2650         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2651
2652         de_buf = dirent_bh->b_data;
2653         limit = de_buf + dir->i_sb->s_blocksize;
2654
2655         while (de_buf < limit) {
2656                 de = (struct ocfs2_dir_entry *)de_buf;
2657
2658                 if (!de->name_len || !de->inode)
2659                         goto inc;
2660
2661                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2662
2663                 trace_ocfs2_dx_dir_index_root_block(
2664                                 (unsigned long long)dir->i_ino,
2665                                 hinfo.major_hash, hinfo.minor_hash,
2666                                 de->name_len, de->name,
2667                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2668
2669                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2670                                            dirent_blk);
2671
2672                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2673 inc:
2674                 de_buf += le16_to_cpu(de->rec_len);
2675         }
2676 }
2677
2678 /*
2679  * Count the number of inline directory entries in di_bh and compare
2680  * them against the number of entries we can hold in an inline dx root
2681  * block.
2682  */
2683 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2684                                          struct buffer_head *di_bh)
2685 {
2686         int dirent_count = 0;
2687         char *de_buf, *limit;
2688         struct ocfs2_dir_entry *de;
2689         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2690
2691         de_buf = di->id2.i_data.id_data;
2692         limit = de_buf + i_size_read(dir);
2693
2694         while (de_buf < limit) {
2695                 de = (struct ocfs2_dir_entry *)de_buf;
2696
2697                 if (de->name_len && de->inode)
2698                         dirent_count++;
2699
2700                 de_buf += le16_to_cpu(de->rec_len);
2701         }
2702
2703         /* We are careful to leave room for one extra record. */
2704         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2705 }
2706
2707 /*
2708  * Expand rec_len of the rightmost dirent in a directory block so that it
2709  * contains the end of our valid space for dirents. We do this during
2710  * expansion from an inline directory to one with extents. The first dir block
2711  * in that case is taken from the inline data portion of the inode block.
2712  *
2713  * This will also return the largest amount of contiguous space for a dirent
2714  * in the block. That value is *not* necessarily the last dirent, even after
2715  * expansion. The directory indexing code wants this value for free space
2716  * accounting. We do this here since we're already walking the entire dir
2717  * block.
2718  *
2719  * We add the dir trailer if this filesystem wants it.
2720  */
2721 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2722                                              struct inode *dir)
2723 {
2724         struct super_block *sb = dir->i_sb;
2725         struct ocfs2_dir_entry *de;
2726         struct ocfs2_dir_entry *prev_de;
2727         char *de_buf, *limit;
2728         unsigned int new_size = sb->s_blocksize;
2729         unsigned int bytes, this_hole;
2730         unsigned int largest_hole = 0;
2731
2732         if (ocfs2_new_dir_wants_trailer(dir))
2733                 new_size = ocfs2_dir_trailer_blk_off(sb);
2734
2735         bytes = new_size - old_size;
2736
2737         limit = start + old_size;
2738         de_buf = start;
2739         de = (struct ocfs2_dir_entry *)de_buf;
2740         do {
2741                 this_hole = ocfs2_figure_dirent_hole(de);
2742                 if (this_hole > largest_hole)
2743                         largest_hole = this_hole;
2744
2745                 prev_de = de;
2746                 de_buf += le16_to_cpu(de->rec_len);
2747                 de = (struct ocfs2_dir_entry *)de_buf;
2748         } while (de_buf < limit);
2749
2750         le16_add_cpu(&prev_de->rec_len, bytes);
2751
2752         /* We need to double check this after modification of the final
2753          * dirent. */
2754         this_hole = ocfs2_figure_dirent_hole(prev_de);
2755         if (this_hole > largest_hole)
2756                 largest_hole = this_hole;
2757
2758         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2759                 return largest_hole;
2760         return 0;
2761 }
2762
2763 /*
2764  * We allocate enough clusters to fulfill "blocks_wanted", but set
2765  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2766  * rest automatically for us.
2767  *
2768  * *first_block_bh is a pointer to the 1st data block allocated to the
2769  *  directory.
2770  */
2771 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2772                                    unsigned int blocks_wanted,
2773                                    struct ocfs2_dir_lookup_result *lookup,
2774                                    struct buffer_head **first_block_bh)
2775 {
2776         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2777         struct super_block *sb = dir->i_sb;
2778         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2779                 credits = ocfs2_inline_to_extents_credits(sb);
2780         u64 dx_insert_blkno, blkno,
2781                 bytes = blocks_wanted << sb->s_blocksize_bits;
2782         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2783         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2784         struct ocfs2_alloc_context *data_ac = NULL;
2785         struct ocfs2_alloc_context *meta_ac = NULL;
2786         struct buffer_head *dirdata_bh = NULL;
2787         struct buffer_head *dx_root_bh = NULL;
2788         struct buffer_head **dx_leaves = NULL;
2789         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2790         handle_t *handle;
2791         struct ocfs2_extent_tree et;
2792         struct ocfs2_extent_tree dx_et;
2793         int did_quota = 0, bytes_allocated = 0;
2794
2795         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2796
2797         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2798         dx_alloc = 0;
2799
2800         down_write(&oi->ip_alloc_sem);
2801
2802         if (ocfs2_supports_indexed_dirs(osb)) {
2803                 credits += ocfs2_add_dir_index_credits(sb);
2804
2805                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2806                 if (!dx_inline) {
2807                         /* Add one more cluster for an index leaf */
2808                         dx_alloc++;
2809                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2810                                                                 &num_dx_leaves);
2811                         if (!dx_leaves) {
2812                                 ret = -ENOMEM;
2813                                 mlog_errno(ret);
2814                                 goto out;
2815                         }
2816                 }
2817
2818                 /* This gets us the dx_root */
2819                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2820                 if (ret) {
2821                         mlog_errno(ret);
2822                         goto out;
2823                 }
2824         }
2825
2826         /*
2827          * We should never need more than 2 clusters for the unindexed
2828          * tree - maximum dirent size is far less than one block. In
2829          * fact, the only time we'd need more than one cluster is if
2830          * blocksize == clustersize and the dirent won't fit in the
2831          * extra space that the expansion to a single block gives. As
2832          * of today, that only happens on 4k/4k file systems.
2833          */
2834         BUG_ON(alloc > 2);
2835
2836         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2837         if (ret) {
2838                 mlog_errno(ret);
2839                 goto out;
2840         }
2841
2842         /*
2843          * Prepare for worst case allocation scenario of two separate
2844          * extents in the unindexed tree.
2845          */
2846         if (alloc == 2)
2847                 credits += OCFS2_SUBALLOC_ALLOC;
2848
2849         handle = ocfs2_start_trans(osb, credits);
2850         if (IS_ERR(handle)) {
2851                 ret = PTR_ERR(handle);
2852                 mlog_errno(ret);
2853                 goto out;
2854         }
2855
2856         ret = dquot_alloc_space_nodirty(dir,
2857                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2858         if (ret)
2859                 goto out_commit;
2860         did_quota = 1;
2861
2862         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2863                 /*
2864                  * Allocate our index cluster first, to maximize the
2865                  * possibility that unindexed leaves grow
2866                  * contiguously.
2867                  */
2868                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2869                                                  dx_leaves, num_dx_leaves,
2870                                                  &dx_insert_blkno);
2871                 if (ret) {
2872                         mlog_errno(ret);
2873                         goto out_commit;
2874                 }
2875                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2876         }
2877
2878         /*
2879          * Try to claim as many clusters as the bitmap can give though
2880          * if we only get one now, that's enough to continue. The rest
2881          * will be claimed after the conversion to extents.
2882          */
2883         if (ocfs2_dir_resv_allowed(osb))
2884                 data_ac->ac_resv = &oi->ip_la_data_resv;
2885         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2886         if (ret) {
2887                 mlog_errno(ret);
2888                 goto out_commit;
2889         }
2890         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2891
2892         /*
2893          * Operations are carefully ordered so that we set up the new
2894          * data block first. The conversion from inline data to
2895          * extents follows.
2896          */
2897         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2898         dirdata_bh = sb_getblk(sb, blkno);
2899         if (!dirdata_bh) {
2900                 ret = -ENOMEM;
2901                 mlog_errno(ret);
2902                 goto out_commit;
2903         }
2904
2905         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2906
2907         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2908                                       OCFS2_JOURNAL_ACCESS_CREATE);
2909         if (ret) {
2910                 mlog_errno(ret);
2911                 goto out_commit;
2912         }
2913
2914         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2915         memset(dirdata_bh->b_data + i_size_read(dir), 0,
2916                sb->s_blocksize - i_size_read(dir));
2917         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2918         if (ocfs2_new_dir_wants_trailer(dir)) {
2919                 /*
2920                  * Prepare the dir trailer up front. It will otherwise look
2921                  * like a valid dirent. Even if inserting the index fails
2922                  * (unlikely), then all we'll have done is given first dir
2923                  * block a small amount of fragmentation.
2924                  */
2925                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2926         }
2927
2928         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2929         ocfs2_journal_dirty(handle, dirdata_bh);
2930
2931         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2932                 /*
2933                  * Dx dirs with an external cluster need to do this up
2934                  * front. Inline dx root's get handled later, after
2935                  * we've allocated our root block. We get passed back
2936                  * a total number of items so that dr_num_entries can
2937                  * be correctly set once the dx_root has been
2938                  * allocated.
2939                  */
2940                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2941                                                num_dx_leaves, &num_dx_entries,
2942                                                dirdata_bh);
2943                 if (ret) {
2944                         mlog_errno(ret);
2945                         goto out_commit;
2946                 }
2947         }
2948
2949         /*
2950          * Set extent, i_size, etc on the directory. After this, the
2951          * inode should contain the same exact dirents as before and
2952          * be fully accessible from system calls.
2953          *
2954          * We let the later dirent insert modify c/mtime - to the user
2955          * the data hasn't changed.
2956          */
2957         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2958                                       OCFS2_JOURNAL_ACCESS_CREATE);
2959         if (ret) {
2960                 mlog_errno(ret);
2961                 goto out_commit;
2962         }
2963
2964         spin_lock(&oi->ip_lock);
2965         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
2966         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
2967         spin_unlock(&oi->ip_lock);
2968
2969         ocfs2_dinode_new_extent_list(dir, di);
2970
2971         i_size_write(dir, sb->s_blocksize);
2972         inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
2973
2974         di->i_size = cpu_to_le64(sb->s_blocksize);
2975         di->i_ctime = di->i_mtime = cpu_to_le64(inode_get_ctime_sec(dir));
2976         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode_get_ctime_nsec(dir));
2977         ocfs2_update_inode_fsync_trans(handle, dir, 1);
2978
2979         /*
2980          * This should never fail as our extent list is empty and all
2981          * related blocks have been journaled already.
2982          */
2983         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
2984                                   0, NULL);
2985         if (ret) {
2986                 mlog_errno(ret);
2987                 goto out_commit;
2988         }
2989
2990         /*
2991          * Set i_blocks after the extent insert for the most up to
2992          * date ip_clusters value.
2993          */
2994         dir->i_blocks = ocfs2_inode_sector_count(dir);
2995
2996         ocfs2_journal_dirty(handle, di_bh);
2997
2998         if (ocfs2_supports_indexed_dirs(osb)) {
2999                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3000                                                 dirdata_bh, meta_ac, dx_inline,
3001                                                 num_dx_entries, &dx_root_bh);
3002                 if (ret) {
3003                         mlog_errno(ret);
3004                         goto out_commit;
3005                 }
3006
3007                 if (dx_inline) {
3008                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3009                                                       dirdata_bh);
3010                 } else {
3011                         ocfs2_init_dx_root_extent_tree(&dx_et,
3012                                                        INODE_CACHE(dir),
3013                                                        dx_root_bh);
3014                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3015                                                   dx_insert_blkno, 1, 0, NULL);
3016                         if (ret)
3017                                 mlog_errno(ret);
3018                 }
3019         }
3020
3021         /*
3022          * We asked for two clusters, but only got one in the 1st
3023          * pass. Claim the 2nd cluster as a separate extent.
3024          */
3025         if (alloc > len) {
3026                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3027                                            &len);
3028                 if (ret) {
3029                         mlog_errno(ret);
3030                         goto out_commit;
3031                 }
3032                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3033
3034                 ret = ocfs2_insert_extent(handle, &et, 1,
3035                                           blkno, len, 0, NULL);
3036                 if (ret) {
3037                         mlog_errno(ret);
3038                         goto out_commit;
3039                 }
3040                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3041         }
3042
3043         *first_block_bh = dirdata_bh;
3044         dirdata_bh = NULL;
3045         if (ocfs2_supports_indexed_dirs(osb)) {
3046                 unsigned int off;
3047
3048                 if (!dx_inline) {
3049                         /*
3050                          * We need to return the correct block within the
3051                          * cluster which should hold our entry.
3052                          */
3053                         off = ocfs2_dx_dir_hash_idx(osb,
3054                                                     &lookup->dl_hinfo);
3055                         get_bh(dx_leaves[off]);
3056                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3057                 }
3058                 lookup->dl_dx_root_bh = dx_root_bh;
3059                 dx_root_bh = NULL;
3060         }
3061
3062 out_commit:
3063         if (ret < 0 && did_quota)
3064                 dquot_free_space_nodirty(dir, bytes_allocated);
3065
3066         ocfs2_commit_trans(osb, handle);
3067
3068 out:
3069         up_write(&oi->ip_alloc_sem);
3070         if (data_ac)
3071                 ocfs2_free_alloc_context(data_ac);
3072         if (meta_ac)
3073                 ocfs2_free_alloc_context(meta_ac);
3074
3075         if (dx_leaves) {
3076                 for (i = 0; i < num_dx_leaves; i++)
3077                         brelse(dx_leaves[i]);
3078                 kfree(dx_leaves);
3079         }
3080
3081         brelse(dirdata_bh);
3082         brelse(dx_root_bh);
3083
3084         return ret;
3085 }
3086
3087 /* returns a bh of the 1st new block in the allocation. */
3088 static int ocfs2_do_extend_dir(struct super_block *sb,
3089                                handle_t *handle,
3090                                struct inode *dir,
3091                                struct buffer_head *parent_fe_bh,
3092                                struct ocfs2_alloc_context *data_ac,
3093                                struct ocfs2_alloc_context *meta_ac,
3094                                struct buffer_head **new_bh)
3095 {
3096         int status;
3097         int extend, did_quota = 0;
3098         u64 p_blkno, v_blkno;
3099
3100         spin_lock(&OCFS2_I(dir)->ip_lock);
3101         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3102         spin_unlock(&OCFS2_I(dir)->ip_lock);
3103
3104         if (extend) {
3105                 u32 offset = OCFS2_I(dir)->ip_clusters;
3106
3107                 status = dquot_alloc_space_nodirty(dir,
3108                                         ocfs2_clusters_to_bytes(sb, 1));
3109                 if (status)
3110                         goto bail;
3111                 did_quota = 1;
3112
3113                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3114                                               1, 0, parent_fe_bh, handle,
3115                                               data_ac, meta_ac, NULL);
3116                 BUG_ON(status == -EAGAIN);
3117                 if (status < 0) {
3118                         mlog_errno(status);
3119                         goto bail;
3120                 }
3121         }
3122
3123         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3124         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3125         if (status < 0) {
3126                 mlog_errno(status);
3127                 goto bail;
3128         }
3129
3130         *new_bh = sb_getblk(sb, p_blkno);
3131         if (!*new_bh) {
3132                 status = -ENOMEM;
3133                 mlog_errno(status);
3134                 goto bail;
3135         }
3136         status = 0;
3137 bail:
3138         if (did_quota && status < 0)
3139                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3140         return status;
3141 }
3142
3143 /*
3144  * Assumes you already have a cluster lock on the directory.
3145  *
3146  * 'blocks_wanted' is only used if we have an inline directory which
3147  * is to be turned into an extent based one. The size of the dirent to
3148  * insert might be larger than the space gained by growing to just one
3149  * block, so we may have to grow the inode by two blocks in that case.
3150  *
3151  * If the directory is already indexed, dx_root_bh must be provided.
3152  */
3153 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3154                             struct inode *dir,
3155                             struct buffer_head *parent_fe_bh,
3156                             unsigned int blocks_wanted,
3157                             struct ocfs2_dir_lookup_result *lookup,
3158                             struct buffer_head **new_de_bh)
3159 {
3160         int status = 0;
3161         int credits, num_free_extents, drop_alloc_sem = 0;
3162         loff_t dir_i_size;
3163         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3164         struct ocfs2_extent_list *el = &fe->id2.i_list;
3165         struct ocfs2_alloc_context *data_ac = NULL;
3166         struct ocfs2_alloc_context *meta_ac = NULL;
3167         handle_t *handle = NULL;
3168         struct buffer_head *new_bh = NULL;
3169         struct ocfs2_dir_entry * de;
3170         struct super_block *sb = osb->sb;
3171         struct ocfs2_extent_tree et;
3172         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3173
3174         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3175                 /*
3176                  * This would be a code error as an inline directory should
3177                  * never have an index root.
3178                  */
3179                 BUG_ON(dx_root_bh);
3180
3181                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3182                                                  blocks_wanted, lookup,
3183                                                  &new_bh);
3184                 if (status) {
3185                         mlog_errno(status);
3186                         goto bail;
3187                 }
3188
3189                 /* Expansion from inline to an indexed directory will
3190                  * have given us this. */
3191                 dx_root_bh = lookup->dl_dx_root_bh;
3192
3193                 if (blocks_wanted == 1) {
3194                         /*
3195                          * If the new dirent will fit inside the space
3196                          * created by pushing out to one block, then
3197                          * we can complete the operation
3198                          * here. Otherwise we have to expand i_size
3199                          * and format the 2nd block below.
3200                          */
3201                         BUG_ON(new_bh == NULL);
3202                         goto bail_bh;
3203                 }
3204
3205                 /*
3206                  * Get rid of 'new_bh' - we want to format the 2nd
3207                  * data block and return that instead.
3208                  */
3209                 brelse(new_bh);
3210                 new_bh = NULL;
3211
3212                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3213                 drop_alloc_sem = 1;
3214                 dir_i_size = i_size_read(dir);
3215                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3216                 goto do_extend;
3217         }
3218
3219         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3220         drop_alloc_sem = 1;
3221         dir_i_size = i_size_read(dir);
3222         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3223                                dir_i_size);
3224
3225         /* dir->i_size is always block aligned. */
3226         spin_lock(&OCFS2_I(dir)->ip_lock);
3227         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3228                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3229                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3230                                               parent_fe_bh);
3231                 num_free_extents = ocfs2_num_free_extents(&et);
3232                 if (num_free_extents < 0) {
3233                         status = num_free_extents;
3234                         mlog_errno(status);
3235                         goto bail;
3236                 }
3237
3238                 if (!num_free_extents) {
3239                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3240                         if (status < 0) {
3241                                 if (status != -ENOSPC)
3242                                         mlog_errno(status);
3243                                 goto bail;
3244                         }
3245                 }
3246
3247                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3248                 if (status < 0) {
3249                         if (status != -ENOSPC)
3250                                 mlog_errno(status);
3251                         goto bail;
3252                 }
3253
3254                 if (ocfs2_dir_resv_allowed(osb))
3255                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3256
3257                 credits = ocfs2_calc_extend_credits(sb, el);
3258         } else {
3259                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3260                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3261         }
3262
3263 do_extend:
3264         if (ocfs2_dir_indexed(dir))
3265                 credits++; /* For attaching the new dirent block to the
3266                             * dx_root */
3267
3268         handle = ocfs2_start_trans(osb, credits);
3269         if (IS_ERR(handle)) {
3270                 status = PTR_ERR(handle);
3271                 handle = NULL;
3272                 mlog_errno(status);
3273                 goto bail;
3274         }
3275
3276         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3277                                      data_ac, meta_ac, &new_bh);
3278         if (status < 0) {
3279                 mlog_errno(status);
3280                 goto bail;
3281         }
3282
3283         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3284
3285         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3286                                          OCFS2_JOURNAL_ACCESS_CREATE);
3287         if (status < 0) {
3288                 mlog_errno(status);
3289                 goto bail;
3290         }
3291         memset(new_bh->b_data, 0, sb->s_blocksize);
3292
3293         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3294         de->inode = 0;
3295         if (ocfs2_supports_dir_trailer(dir)) {
3296                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3297
3298                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3299
3300                 if (ocfs2_dir_indexed(dir)) {
3301                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3302                                                            dx_root_bh, new_bh);
3303                         if (status) {
3304                                 mlog_errno(status);
3305                                 goto bail;
3306                         }
3307                 }
3308         } else {
3309                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3310         }
3311         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3312         ocfs2_journal_dirty(handle, new_bh);
3313
3314         dir_i_size += dir->i_sb->s_blocksize;
3315         i_size_write(dir, dir_i_size);
3316         dir->i_blocks = ocfs2_inode_sector_count(dir);
3317         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3318         if (status < 0) {
3319                 mlog_errno(status);
3320                 goto bail;
3321         }
3322
3323 bail_bh:
3324         *new_de_bh = new_bh;
3325         get_bh(*new_de_bh);
3326 bail:
3327         if (handle)
3328                 ocfs2_commit_trans(osb, handle);
3329         if (drop_alloc_sem)
3330                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3331
3332         if (data_ac)
3333                 ocfs2_free_alloc_context(data_ac);
3334         if (meta_ac)
3335                 ocfs2_free_alloc_context(meta_ac);
3336
3337         brelse(new_bh);
3338
3339         return status;
3340 }
3341
3342 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3343                                    const char *name, int namelen,
3344                                    struct buffer_head **ret_de_bh,
3345                                    unsigned int *blocks_wanted)
3346 {
3347         int ret;
3348         struct super_block *sb = dir->i_sb;
3349         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3350         struct ocfs2_dir_entry *de, *last_de = NULL;
3351         char *first_de, *de_buf, *limit;
3352         unsigned long offset = 0;
3353         unsigned int rec_len, new_rec_len, free_space;
3354
3355         /*
3356          * This calculates how many free bytes we'd have in block zero, should
3357          * this function force expansion to an extent tree.
3358          */
3359         if (ocfs2_new_dir_wants_trailer(dir))
3360                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3361         else
3362                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3363
3364         first_de = di->id2.i_data.id_data;
3365         de_buf = first_de;
3366         limit = de_buf + i_size_read(dir);
3367         rec_len = OCFS2_DIR_REC_LEN(namelen);
3368
3369         while (de_buf < limit) {
3370                 de = (struct ocfs2_dir_entry *)de_buf;
3371
3372                 if (!ocfs2_check_dir_entry(dir, de, di_bh, first_de,
3373                                            i_size_read(dir), offset)) {
3374                         ret = -ENOENT;
3375                         goto out;
3376                 }
3377                 if (ocfs2_match(namelen, name, de)) {
3378                         ret = -EEXIST;
3379                         goto out;
3380                 }
3381                 /*
3382                  * No need to check for a trailing dirent record here as
3383                  * they're not used for inline dirs.
3384                  */
3385
3386                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3387                         /* Ok, we found a spot. Return this bh and let
3388                          * the caller actually fill it in. */
3389                         *ret_de_bh = di_bh;
3390                         get_bh(*ret_de_bh);
3391                         ret = 0;
3392                         goto out;
3393                 }
3394
3395                 last_de = de;
3396                 de_buf += le16_to_cpu(de->rec_len);
3397                 offset += le16_to_cpu(de->rec_len);
3398         }
3399
3400         /*
3401          * We're going to require expansion of the directory - figure
3402          * out how many blocks we'll need so that a place for the
3403          * dirent can be found.
3404          */
3405         *blocks_wanted = 1;
3406         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3407         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3408                 *blocks_wanted = 2;
3409
3410         ret = -ENOSPC;
3411 out:
3412         return ret;
3413 }
3414
3415 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3416                                    int namelen, struct buffer_head **ret_de_bh)
3417 {
3418         unsigned long offset;
3419         struct buffer_head *bh = NULL;
3420         unsigned short rec_len;
3421         struct ocfs2_dir_entry *de;
3422         struct super_block *sb = dir->i_sb;
3423         int status;
3424         int blocksize = dir->i_sb->s_blocksize;
3425
3426         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3427         if (status)
3428                 goto bail;
3429
3430         rec_len = OCFS2_DIR_REC_LEN(namelen);
3431         offset = 0;
3432         de = (struct ocfs2_dir_entry *) bh->b_data;
3433         while (1) {
3434                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3435                         brelse(bh);
3436                         bh = NULL;
3437
3438                         if (i_size_read(dir) <= offset) {
3439                                 /*
3440                                  * Caller will have to expand this
3441                                  * directory.
3442                                  */
3443                                 status = -ENOSPC;
3444                                 goto bail;
3445                         }
3446                         status = ocfs2_read_dir_block(dir,
3447                                              offset >> sb->s_blocksize_bits,
3448                                              &bh, 0);
3449                         if (status)
3450                                 goto bail;
3451
3452                         /* move to next block */
3453                         de = (struct ocfs2_dir_entry *) bh->b_data;
3454                 }
3455                 if (!ocfs2_check_dir_entry(dir, de, bh, bh->b_data, blocksize,
3456                                            offset)) {
3457                         status = -ENOENT;
3458                         goto bail;
3459                 }
3460                 if (ocfs2_match(namelen, name, de)) {
3461                         status = -EEXIST;
3462                         goto bail;
3463                 }
3464
3465                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3466                                            blocksize))
3467                         goto next;
3468
3469                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3470                         /* Ok, we found a spot. Return this bh and let
3471                          * the caller actually fill it in. */
3472                         *ret_de_bh = bh;
3473                         get_bh(*ret_de_bh);
3474                         status = 0;
3475                         goto bail;
3476                 }
3477 next:
3478                 offset += le16_to_cpu(de->rec_len);
3479                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3480         }
3481
3482 bail:
3483         brelse(bh);
3484         if (status)
3485                 mlog_errno(status);
3486
3487         return status;
3488 }
3489
3490 static int dx_leaf_sort_cmp(const void *a, const void *b)
3491 {
3492         const struct ocfs2_dx_entry *entry1 = a;
3493         const struct ocfs2_dx_entry *entry2 = b;
3494         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3495         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3496         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3497         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3498
3499         if (major_hash1 > major_hash2)
3500                 return 1;
3501         if (major_hash1 < major_hash2)
3502                 return -1;
3503
3504         /*
3505          * It is not strictly necessary to sort by minor
3506          */
3507         if (minor_hash1 > minor_hash2)
3508                 return 1;
3509         if (minor_hash1 < minor_hash2)
3510                 return -1;
3511         return 0;
3512 }
3513
3514 static void dx_leaf_sort_swap(void *a, void *b, int size)
3515 {
3516         struct ocfs2_dx_entry *entry1 = a;
3517         struct ocfs2_dx_entry *entry2 = b;
3518
3519         BUG_ON(size != sizeof(*entry1));
3520
3521         swap(*entry1, *entry2);
3522 }
3523
3524 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3525 {
3526         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3527         int i, num = le16_to_cpu(dl_list->de_num_used);
3528
3529         for (i = 0; i < (num - 1); i++) {
3530                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3531                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3532                         return 0;
3533         }
3534
3535         return 1;
3536 }
3537
3538 /*
3539  * Find the optimal value to split this leaf on. This expects the leaf
3540  * entries to be in sorted order.
3541  *
3542  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3543  * the hash we want to insert.
3544  *
3545  * This function is only concerned with the major hash - that which
3546  * determines which cluster an item belongs to.
3547  */
3548 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3549                                         u32 leaf_cpos, u32 insert_hash,
3550                                         u32 *split_hash)
3551 {
3552         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3553         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3554         int allsame;
3555
3556         /*
3557          * There's a couple rare, but nasty corner cases we have to
3558          * check for here. All of them involve a leaf where all value
3559          * have the same hash, which is what we look for first.
3560          *
3561          * Most of the time, all of the above is false, and we simply
3562          * pick the median value for a split.
3563          */
3564         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3565         if (allsame) {
3566                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3567
3568                 if (val == insert_hash) {
3569                         /*
3570                          * No matter where we would choose to split,
3571                          * the new entry would want to occupy the same
3572                          * block as these. Since there's no space left
3573                          * in their existing block, we know there
3574                          * won't be space after the split.
3575                          */
3576                         return -ENOSPC;
3577                 }
3578
3579                 if (val == leaf_cpos) {
3580                         /*
3581                          * Because val is the same as leaf_cpos (which
3582                          * is the smallest value this leaf can have),
3583                          * yet is not equal to insert_hash, then we
3584                          * know that insert_hash *must* be larger than
3585                          * val (and leaf_cpos). At least cpos+1 in value.
3586                          *
3587                          * We also know then, that there cannot be an
3588                          * adjacent extent (otherwise we'd be looking
3589                          * at it). Choosing this value gives us a
3590                          * chance to get some contiguousness.
3591                          */
3592                         *split_hash = leaf_cpos + 1;
3593                         return 0;
3594                 }
3595
3596                 if (val > insert_hash) {
3597                         /*
3598                          * val can not be the same as insert hash, and
3599                          * also must be larger than leaf_cpos. Also,
3600                          * we know that there can't be a leaf between
3601                          * cpos and val, otherwise the entries with
3602                          * hash 'val' would be there.
3603                          */
3604                         *split_hash = val;
3605                         return 0;
3606                 }
3607
3608                 *split_hash = insert_hash;
3609                 return 0;
3610         }
3611
3612         /*
3613          * Since the records are sorted and the checks above
3614          * guaranteed that not all records in this block are the same,
3615          * we simple travel forward, from the median, and pick the 1st
3616          * record whose value is larger than leaf_cpos.
3617          */
3618         for (i = (num_used / 2); i < num_used; i++)
3619                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3620                     leaf_cpos)
3621                         break;
3622
3623         BUG_ON(i == num_used); /* Should be impossible */
3624         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3625         return 0;
3626 }
3627
3628 /*
3629  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3630  * larger than split_hash into new_dx_leaves. We use a temporary
3631  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3632  *
3633  * Since the block offset inside a leaf (cluster) is a constant mask
3634  * of minor_hash, we can optimize - an item at block offset X within
3635  * the original cluster, will be at offset X within the new cluster.
3636  */
3637 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3638                                        handle_t *handle,
3639                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3640                                        struct buffer_head **orig_dx_leaves,
3641                                        struct buffer_head **new_dx_leaves,
3642                                        int num_dx_leaves)
3643 {
3644         int i, j, num_used;
3645         u32 major_hash;
3646         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3647         struct ocfs2_dx_entry_list *orig_list, *tmp_list;
3648         struct ocfs2_dx_entry *dx_entry;
3649
3650         tmp_list = &tmp_dx_leaf->dl_list;
3651
3652         for (i = 0; i < num_dx_leaves; i++) {
3653                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3654                 orig_list = &orig_dx_leaf->dl_list;
3655                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3656
3657                 num_used = le16_to_cpu(orig_list->de_num_used);
3658
3659                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3660                 tmp_list->de_num_used = cpu_to_le16(0);
3661                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3662
3663                 for (j = 0; j < num_used; j++) {
3664                         dx_entry = &orig_list->de_entries[j];
3665                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3666                         if (major_hash >= split_hash)
3667                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3668                                                               dx_entry);
3669                         else
3670                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3671                                                               dx_entry);
3672                 }
3673                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3674
3675                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3676                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3677         }
3678 }
3679
3680 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3681                                           struct ocfs2_dx_root_block *dx_root)
3682 {
3683         int credits = ocfs2_clusters_to_blocks(osb->sb, 3);
3684
3685         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3686         credits += ocfs2_quota_trans_credits(osb->sb);
3687         return credits;
3688 }
3689
3690 /*
3691  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3692  * half our entries into.
3693  */
3694 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3695                                   struct buffer_head *dx_root_bh,
3696                                   struct buffer_head *dx_leaf_bh,
3697                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3698                                   u64 leaf_blkno)
3699 {
3700         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3701         int credits, ret, i, num_used, did_quota = 0;
3702         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3703         u64 orig_leaves_start;
3704         int num_dx_leaves;
3705         struct buffer_head **orig_dx_leaves = NULL;
3706         struct buffer_head **new_dx_leaves = NULL;
3707         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3708         struct ocfs2_extent_tree et;
3709         handle_t *handle = NULL;
3710         struct ocfs2_dx_root_block *dx_root;
3711         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3712
3713         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3714                                      (unsigned long long)leaf_blkno,
3715                                      insert_hash);
3716
3717         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3718
3719         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3720         /*
3721          * XXX: This is a rather large limit. We should use a more
3722          * realistic value.
3723          */
3724         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3725                 return -ENOSPC;
3726
3727         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3728         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3729                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3730                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3731                      (unsigned long long)leaf_blkno, num_used);
3732                 ret = -EIO;
3733                 goto out;
3734         }
3735
3736         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3737         if (!orig_dx_leaves) {
3738                 ret = -ENOMEM;
3739                 mlog_errno(ret);
3740                 goto out;
3741         }
3742
3743         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3744         if (!new_dx_leaves) {
3745                 ret = -ENOMEM;
3746                 mlog_errno(ret);
3747                 goto out;
3748         }
3749
3750         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3751         if (ret) {
3752                 if (ret != -ENOSPC)
3753                         mlog_errno(ret);
3754                 goto out;
3755         }
3756
3757         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3758         handle = ocfs2_start_trans(osb, credits);
3759         if (IS_ERR(handle)) {
3760                 ret = PTR_ERR(handle);
3761                 handle = NULL;
3762                 mlog_errno(ret);
3763                 goto out;
3764         }
3765
3766         ret = dquot_alloc_space_nodirty(dir,
3767                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3768         if (ret)
3769                 goto out_commit;
3770         did_quota = 1;
3771
3772         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3773                                       OCFS2_JOURNAL_ACCESS_WRITE);
3774         if (ret) {
3775                 mlog_errno(ret);
3776                 goto out_commit;
3777         }
3778
3779         /*
3780          * This block is changing anyway, so we can sort it in place.
3781          */
3782         sort(dx_leaf->dl_list.de_entries, num_used,
3783              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3784              dx_leaf_sort_swap);
3785
3786         ocfs2_journal_dirty(handle, dx_leaf_bh);
3787
3788         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3789                                            &split_hash);
3790         if (ret) {
3791                 mlog_errno(ret);
3792                 goto  out_commit;
3793         }
3794
3795         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3796
3797         /*
3798          * We have to carefully order operations here. There are items
3799          * which want to be in the new cluster before insert, but in
3800          * order to put those items in the new cluster, we alter the
3801          * old cluster. A failure to insert gets nasty.
3802          *
3803          * So, start by reserving writes to the old
3804          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3805          * the new cluster for us, before inserting it. The insert
3806          * won't happen if there's an error before that. Once the
3807          * insert is done then, we can transfer from one leaf into the
3808          * other without fear of hitting any error.
3809          */
3810
3811         /*
3812          * The leaf transfer wants some scratch space so that we don't
3813          * wind up doing a bunch of expensive memmove().
3814          */
3815         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3816         if (!tmp_dx_leaf) {
3817                 ret = -ENOMEM;
3818                 mlog_errno(ret);
3819                 goto out_commit;
3820         }
3821
3822         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3823         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3824                                    orig_dx_leaves);
3825         if (ret) {
3826                 mlog_errno(ret);
3827                 goto out_commit;
3828         }
3829
3830         cpos = split_hash;
3831         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3832                                        data_ac, meta_ac, new_dx_leaves,
3833                                        num_dx_leaves);
3834         if (ret) {
3835                 mlog_errno(ret);
3836                 goto out_commit;
3837         }
3838
3839         for (i = 0; i < num_dx_leaves; i++) {
3840                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3841                                               orig_dx_leaves[i],
3842                                               OCFS2_JOURNAL_ACCESS_WRITE);
3843                 if (ret) {
3844                         mlog_errno(ret);
3845                         goto out_commit;
3846                 }
3847
3848                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3849                                               new_dx_leaves[i],
3850                                               OCFS2_JOURNAL_ACCESS_WRITE);
3851                 if (ret) {
3852                         mlog_errno(ret);
3853                         goto out_commit;
3854                 }
3855         }
3856
3857         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3858                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3859
3860 out_commit:
3861         if (ret < 0 && did_quota)
3862                 dquot_free_space_nodirty(dir,
3863                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3864
3865         ocfs2_update_inode_fsync_trans(handle, dir, 1);
3866         ocfs2_commit_trans(osb, handle);
3867
3868 out:
3869         if (orig_dx_leaves || new_dx_leaves) {
3870                 for (i = 0; i < num_dx_leaves; i++) {
3871                         if (orig_dx_leaves)
3872                                 brelse(orig_dx_leaves[i]);
3873                         if (new_dx_leaves)
3874                                 brelse(new_dx_leaves[i]);
3875                 }
3876                 kfree(orig_dx_leaves);
3877                 kfree(new_dx_leaves);
3878         }
3879
3880         if (meta_ac)
3881                 ocfs2_free_alloc_context(meta_ac);
3882         if (data_ac)
3883                 ocfs2_free_alloc_context(data_ac);
3884
3885         kfree(tmp_dx_leaf);
3886         return ret;
3887 }
3888
3889 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3890                                    struct buffer_head *di_bh,
3891                                    struct buffer_head *dx_root_bh,
3892                                    const char *name, int namelen,
3893                                    struct ocfs2_dir_lookup_result *lookup)
3894 {
3895         int ret, rebalanced = 0;
3896         struct ocfs2_dx_root_block *dx_root;
3897         struct buffer_head *dx_leaf_bh = NULL;
3898         struct ocfs2_dx_leaf *dx_leaf;
3899         u64 blkno;
3900         u32 leaf_cpos;
3901
3902         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3903
3904 restart_search:
3905         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3906                                   &leaf_cpos, &blkno);
3907         if (ret) {
3908                 mlog_errno(ret);
3909                 goto out;
3910         }
3911
3912         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3913         if (ret) {
3914                 mlog_errno(ret);
3915                 goto out;
3916         }
3917
3918         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3919
3920         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3921             le16_to_cpu(dx_leaf->dl_list.de_count)) {
3922                 if (rebalanced) {
3923                         /*
3924                          * Rebalancing should have provided us with
3925                          * space in an appropriate leaf.
3926                          *
3927                          * XXX: Is this an abnormal condition then?
3928                          * Should we print a message here?
3929                          */
3930                         ret = -ENOSPC;
3931                         goto out;
3932                 }
3933
3934                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3935                                              &lookup->dl_hinfo, leaf_cpos,
3936                                              blkno);
3937                 if (ret) {
3938                         if (ret != -ENOSPC)
3939                                 mlog_errno(ret);
3940                         goto out;
3941                 }
3942
3943                 /*
3944                  * Restart the lookup. The rebalance might have
3945                  * changed which block our item fits into. Mark our
3946                  * progress, so we only execute this once.
3947                  */
3948                 brelse(dx_leaf_bh);
3949                 dx_leaf_bh = NULL;
3950                 rebalanced = 1;
3951                 goto restart_search;
3952         }
3953
3954         lookup->dl_dx_leaf_bh = dx_leaf_bh;
3955         dx_leaf_bh = NULL;
3956
3957 out:
3958         brelse(dx_leaf_bh);
3959         return ret;
3960 }
3961
3962 static int ocfs2_search_dx_free_list(struct inode *dir,
3963                                      struct buffer_head *dx_root_bh,
3964                                      int namelen,
3965                                      struct ocfs2_dir_lookup_result *lookup)
3966 {
3967         int ret = -ENOSPC;
3968         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
3969         struct ocfs2_dir_block_trailer *db;
3970         u64 next_block;
3971         int rec_len = OCFS2_DIR_REC_LEN(namelen);
3972         struct ocfs2_dx_root_block *dx_root;
3973
3974         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3975         next_block = le64_to_cpu(dx_root->dr_free_blk);
3976
3977         while (next_block) {
3978                 brelse(prev_leaf_bh);
3979                 prev_leaf_bh = leaf_bh;
3980                 leaf_bh = NULL;
3981
3982                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
3983                 if (ret) {
3984                         mlog_errno(ret);
3985                         goto out;
3986                 }
3987
3988                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
3989                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
3990                         lookup->dl_leaf_bh = leaf_bh;
3991                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
3992                         leaf_bh = NULL;
3993                         prev_leaf_bh = NULL;
3994                         break;
3995                 }
3996
3997                 next_block = le64_to_cpu(db->db_free_next);
3998         }
3999
4000         if (!next_block)
4001                 ret = -ENOSPC;
4002
4003 out:
4004
4005         brelse(leaf_bh);
4006         brelse(prev_leaf_bh);
4007         return ret;
4008 }
4009
4010 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4011                                        struct buffer_head *dx_root_bh)
4012 {
4013         int ret, num_dx_leaves, i, j, did_quota = 0;
4014         struct buffer_head **dx_leaves = NULL;
4015         struct ocfs2_extent_tree et;
4016         u64 insert_blkno;
4017         struct ocfs2_alloc_context *data_ac = NULL;
4018         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4019         handle_t *handle = NULL;
4020         struct ocfs2_dx_root_block *dx_root;
4021         struct ocfs2_dx_entry_list *entry_list;
4022         struct ocfs2_dx_entry *dx_entry;
4023         struct ocfs2_dx_leaf *target_leaf;
4024
4025         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4026         if (ret) {
4027                 mlog_errno(ret);
4028                 goto out;
4029         }
4030
4031         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4032         if (!dx_leaves) {
4033                 ret = -ENOMEM;
4034                 mlog_errno(ret);
4035                 goto out;
4036         }
4037
4038         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4039         if (IS_ERR(handle)) {
4040                 ret = PTR_ERR(handle);
4041                 mlog_errno(ret);
4042                 goto out;
4043         }
4044
4045         ret = dquot_alloc_space_nodirty(dir,
4046                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4047         if (ret)
4048                 goto out_commit;
4049         did_quota = 1;
4050
4051         /*
4052          * We do this up front, before the allocation, so that a
4053          * failure to add the dx_root_bh to the journal won't result
4054          * us losing clusters.
4055          */
4056         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4057                                       OCFS2_JOURNAL_ACCESS_WRITE);
4058         if (ret) {
4059                 mlog_errno(ret);
4060                 goto out_commit;
4061         }
4062
4063         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4064                                          num_dx_leaves, &insert_blkno);
4065         if (ret) {
4066                 mlog_errno(ret);
4067                 goto out_commit;
4068         }
4069
4070         /*
4071          * Transfer the entries from our dx_root into the appropriate
4072          * block
4073          */
4074         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4075         entry_list = &dx_root->dr_entries;
4076
4077         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4078                 dx_entry = &entry_list->de_entries[i];
4079
4080                 j = __ocfs2_dx_dir_hash_idx(osb,
4081                                             le32_to_cpu(dx_entry->dx_minor_hash));
4082                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4083
4084                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4085
4086                 /* Each leaf has been passed to the journal already
4087                  * via __ocfs2_dx_dir_new_cluster() */
4088         }
4089
4090         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4091         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4092                offsetof(struct ocfs2_dx_root_block, dr_list));
4093         dx_root->dr_list.l_count =
4094                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4095
4096         /* This should never fail considering we start with an empty
4097          * dx_root. */
4098         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4099         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4100         if (ret)
4101                 mlog_errno(ret);
4102         did_quota = 0;
4103
4104         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4105         ocfs2_journal_dirty(handle, dx_root_bh);
4106
4107 out_commit:
4108         if (ret < 0 && did_quota)
4109                 dquot_free_space_nodirty(dir,
4110                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4111
4112         ocfs2_commit_trans(osb, handle);
4113
4114 out:
4115         if (data_ac)
4116                 ocfs2_free_alloc_context(data_ac);
4117
4118         if (dx_leaves) {
4119                 for (i = 0; i < num_dx_leaves; i++)
4120                         brelse(dx_leaves[i]);
4121                 kfree(dx_leaves);
4122         }
4123         return ret;
4124 }
4125
4126 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4127 {
4128         struct ocfs2_dx_root_block *dx_root;
4129         struct ocfs2_dx_entry_list *entry_list;
4130
4131         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4132         entry_list = &dx_root->dr_entries;
4133
4134         if (le16_to_cpu(entry_list->de_num_used) >=
4135             le16_to_cpu(entry_list->de_count))
4136                 return -ENOSPC;
4137
4138         return 0;
4139 }
4140
4141 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4142                                            struct buffer_head *di_bh,
4143                                            const char *name,
4144                                            int namelen,
4145                                            struct ocfs2_dir_lookup_result *lookup)
4146 {
4147         int ret, free_dx_root = 1;
4148         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4149         struct buffer_head *dx_root_bh = NULL;
4150         struct buffer_head *leaf_bh = NULL;
4151         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4152         struct ocfs2_dx_root_block *dx_root;
4153
4154         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4155         if (ret) {
4156                 mlog_errno(ret);
4157                 goto out;
4158         }
4159
4160         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4161         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4162                 ret = -ENOSPC;
4163                 mlog_errno(ret);
4164                 goto out;
4165         }
4166
4167         if (ocfs2_dx_root_inline(dx_root)) {
4168                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4169
4170                 if (ret == 0)
4171                         goto search_el;
4172
4173                 /*
4174                  * We ran out of room in the root block. Expand it to
4175                  * an extent, then allow ocfs2_find_dir_space_dx to do
4176                  * the rest.
4177                  */
4178                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4179                 if (ret) {
4180                         mlog_errno(ret);
4181                         goto out;
4182                 }
4183         }
4184
4185         /*
4186          * Insert preparation for an indexed directory is split into two
4187          * steps. The call to find_dir_space_dx reserves room in the index for
4188          * an additional item. If we run out of space there, it's a real error
4189          * we can't continue on.
4190          */
4191         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4192                                       namelen, lookup);
4193         if (ret) {
4194                 mlog_errno(ret);
4195                 goto out;
4196         }
4197
4198 search_el:
4199         /*
4200          * Next, we need to find space in the unindexed tree. This call
4201          * searches using the free space linked list. If the unindexed tree
4202          * lacks sufficient space, we'll expand it below. The expansion code
4203          * is smart enough to add any new blocks to the free space list.
4204          */
4205         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4206         if (ret && ret != -ENOSPC) {
4207                 mlog_errno(ret);
4208                 goto out;
4209         }
4210
4211         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4212         lookup->dl_dx_root_bh = dx_root_bh;
4213         free_dx_root = 0;
4214
4215         if (ret == -ENOSPC) {
4216                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4217
4218                 if (ret) {
4219                         mlog_errno(ret);
4220                         goto out;
4221                 }
4222
4223                 /*
4224                  * We make the assumption here that new leaf blocks are added
4225                  * to the front of our free list.
4226                  */
4227                 lookup->dl_prev_leaf_bh = NULL;
4228                 lookup->dl_leaf_bh = leaf_bh;
4229         }
4230
4231 out:
4232         if (free_dx_root)
4233                 brelse(dx_root_bh);
4234         return ret;
4235 }
4236
4237 /*
4238  * Get a directory ready for insert. Any directory allocation required
4239  * happens here. Success returns zero, and enough context in the dir
4240  * lookup result that ocfs2_add_entry() will be able complete the task
4241  * with minimal performance impact.
4242  */
4243 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4244                                  struct inode *dir,
4245                                  struct buffer_head *parent_fe_bh,
4246                                  const char *name,
4247                                  int namelen,
4248                                  struct ocfs2_dir_lookup_result *lookup)
4249 {
4250         int ret;
4251         unsigned int blocks_wanted = 1;
4252         struct buffer_head *bh = NULL;
4253
4254         trace_ocfs2_prepare_dir_for_insert(
4255                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4256
4257         /*
4258          * Do this up front to reduce confusion.
4259          *
4260          * The directory might start inline, then be turned into an
4261          * indexed one, in which case we'd need to hash deep inside
4262          * ocfs2_find_dir_space_id(). Since
4263          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4264          * done, there seems no point in spreading out the calls. We
4265          * can optimize away the case where the file system doesn't
4266          * support indexing.
4267          */
4268         if (ocfs2_supports_indexed_dirs(osb))
4269                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4270
4271         if (ocfs2_dir_indexed(dir)) {
4272                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4273                                                       name, namelen, lookup);
4274                 if (ret)
4275                         mlog_errno(ret);
4276                 goto out;
4277         }
4278
4279         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4280                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4281                                               namelen, &bh, &blocks_wanted);
4282         } else
4283                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4284
4285         if (ret && ret != -ENOSPC) {
4286                 mlog_errno(ret);
4287                 goto out;
4288         }
4289
4290         if (ret == -ENOSPC) {
4291                 /*
4292                  * We have to expand the directory to add this name.
4293                  */
4294                 BUG_ON(bh);
4295
4296                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4297                                        lookup, &bh);
4298                 if (ret) {
4299                         if (ret != -ENOSPC)
4300                                 mlog_errno(ret);
4301                         goto out;
4302                 }
4303
4304                 BUG_ON(!bh);
4305         }
4306
4307         lookup->dl_leaf_bh = bh;
4308         bh = NULL;
4309 out:
4310         brelse(bh);
4311         return ret;
4312 }
4313
4314 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4315                                      struct buffer_head *di_bh,
4316                                      struct buffer_head *dx_root_bh)
4317 {
4318         int ret;
4319         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4320         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4321         struct ocfs2_dx_root_block *dx_root;
4322         struct inode *dx_alloc_inode = NULL;
4323         struct buffer_head *dx_alloc_bh = NULL;
4324         handle_t *handle;
4325         u64 blk;
4326         u16 bit;
4327         u64 bg_blkno;
4328
4329         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4330
4331         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4332                                         EXTENT_ALLOC_SYSTEM_INODE,
4333                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4334         if (!dx_alloc_inode) {
4335                 ret = -ENOMEM;
4336                 mlog_errno(ret);
4337                 goto out;
4338         }
4339         inode_lock(dx_alloc_inode);
4340
4341         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4342         if (ret) {
4343                 mlog_errno(ret);
4344                 goto out_mutex;
4345         }
4346
4347         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4348         if (IS_ERR(handle)) {
4349                 ret = PTR_ERR(handle);
4350                 mlog_errno(ret);
4351                 goto out_unlock;
4352         }
4353
4354         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4355                                       OCFS2_JOURNAL_ACCESS_WRITE);
4356         if (ret) {
4357                 mlog_errno(ret);
4358                 goto out_commit;
4359         }
4360
4361         spin_lock(&OCFS2_I(dir)->ip_lock);
4362         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4363         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4364         spin_unlock(&OCFS2_I(dir)->ip_lock);
4365         di->i_dx_root = cpu_to_le64(0ULL);
4366         ocfs2_update_inode_fsync_trans(handle, dir, 1);
4367
4368         ocfs2_journal_dirty(handle, di_bh);
4369
4370         blk = le64_to_cpu(dx_root->dr_blkno);
4371         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4372         if (dx_root->dr_suballoc_loc)
4373                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4374         else
4375                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4376         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4377                                        bit, bg_blkno, 1);
4378         if (ret)
4379                 mlog_errno(ret);
4380
4381 out_commit:
4382         ocfs2_commit_trans(osb, handle);
4383
4384 out_unlock:
4385         ocfs2_inode_unlock(dx_alloc_inode, 1);
4386
4387 out_mutex:
4388         inode_unlock(dx_alloc_inode);
4389         brelse(dx_alloc_bh);
4390 out:
4391         iput(dx_alloc_inode);
4392         return ret;
4393 }
4394
4395 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4396 {
4397         int ret;
4398         unsigned int clen;
4399         u32 major_hash = UINT_MAX, p_cpos, cpos;
4400         u64 blkno;
4401         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4402         struct buffer_head *dx_root_bh = NULL;
4403         struct ocfs2_dx_root_block *dx_root;
4404         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4405         struct ocfs2_cached_dealloc_ctxt dealloc;
4406         struct ocfs2_extent_tree et;
4407
4408         ocfs2_init_dealloc_ctxt(&dealloc);
4409
4410         if (!ocfs2_dir_indexed(dir))
4411                 return 0;
4412
4413         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4414         if (ret) {
4415                 mlog_errno(ret);
4416                 goto out;
4417         }
4418         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4419
4420         if (ocfs2_dx_root_inline(dx_root))
4421                 goto remove_index;
4422
4423         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4424
4425         /* XXX: What if dr_clusters is too large? */
4426         while (le32_to_cpu(dx_root->dr_clusters)) {
4427                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4428                                               major_hash, &cpos, &blkno, &clen);
4429                 if (ret) {
4430                         mlog_errno(ret);
4431                         goto out;
4432                 }
4433
4434                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4435
4436                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4437                                                &dealloc, 0, false);
4438                 if (ret) {
4439                         mlog_errno(ret);
4440                         goto out;
4441                 }
4442
4443                 if (cpos == 0)
4444                         break;
4445
4446                 major_hash = cpos - 1;
4447         }
4448
4449 remove_index:
4450         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4451         if (ret) {
4452                 mlog_errno(ret);
4453                 goto out;
4454         }
4455
4456         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4457 out:
4458         ocfs2_schedule_truncate_log_flush(osb, 1);
4459         ocfs2_run_deallocs(osb, &dealloc);
4460
4461         brelse(dx_root_bh);
4462         return ret;
4463 }