]> git.ipfire.org Git - thirdparty/xfsprogs-dev.git/blobdiff - libxfs/xfs_dir2_leaf.c
xfs: fix transaction leak on remote attr set/remove failure
[thirdparty/xfsprogs-dev.git] / libxfs / xfs_dir2_leaf.c
index adef6eb0afffa2cd02c774266e7b725f92d0dd92..dbe4d2efb069a6ea6f03c21bf71b723ef2e87b3a 100644 (file)
+// SPDX-License-Identifier: GPL-2.0
 /*
- * Copyright (c) 2000-2002 Silicon Graphics, Inc.  All Rights Reserved.
- *
- * This program is free software; you can redistribute it and/or modify it
- * under the terms of version 2 of the GNU General Public License as
- * published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it would be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
- *
- * Further, this software is distributed without any warranty that it is
- * free of the rightful claim of any third person regarding infringement
- * or the like.  Any license provided herein, whether implied or
- * otherwise, applies only to this software file.  Patent licenses, if
- * any, provided herein do not apply to combinations of this program with
- * other software, or any other product whatsoever.
- *
- * You should have received a copy of the GNU General Public License along
- * with this program; if not, write the Free Software Foundation, Inc., 59
- * Temple Place - Suite 330, Boston MA 02111-1307, USA.
- *
- * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
- * Mountain View, CA  94043, or:
- *
- * http://www.sgi.com
- *
- * For further information regarding this notice, see:
- *
- * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
+ * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
+ * Copyright (c) 2013 Red Hat, Inc.
+ * All Rights Reserved.
  */
+#include "libxfs_priv.h"
+#include "xfs_fs.h"
+#include "xfs_format.h"
+#include "xfs_log_format.h"
+#include "xfs_trans_resv.h"
+#include "xfs_mount.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
+#include "xfs_inode.h"
+#include "xfs_bmap.h"
+#include "xfs_dir2.h"
+#include "xfs_dir2_priv.h"
+#include "xfs_trace.h"
+#include "xfs_trans.h"
+#include "xfs_cksum.h"
 
 /*
- * xfs_dir2_leaf.c
- * XFS directory version 2 implementation - single leaf form
- * see xfs_dir2_leaf.h for data structures.
- * These directories have multiple XFS_DIR2_DATA blocks and one
- * XFS_DIR2_LEAF1 block containing the hash table and freespace map.
+ * Local function declarations.
  */
+static int xfs_dir2_leaf_lookup_int(xfs_da_args_t *args, struct xfs_buf **lbpp,
+                                   int *indexp, struct xfs_buf **dbpp);
+static void xfs_dir3_leaf_log_bests(struct xfs_da_args *args,
+                                   struct xfs_buf *bp, int first, int last);
+static void xfs_dir3_leaf_log_tail(struct xfs_da_args *args,
+                                  struct xfs_buf *bp);
 
-#include <xfs.h>
+/*
+ * Check the internal consistency of a leaf1 block.
+ * Pop an assert if something is wrong.
+ */
+#ifdef DEBUG
+static xfs_failaddr_t
+xfs_dir3_leaf1_check(
+       struct xfs_inode        *dp,
+       struct xfs_buf          *bp)
+{
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+
+       if (leafhdr.magic == XFS_DIR3_LEAF1_MAGIC) {
+               struct xfs_dir3_leaf_hdr *leaf3 = bp->b_addr;
+               if (be64_to_cpu(leaf3->info.blkno) != bp->b_bn)
+                       return __this_address;
+       } else if (leafhdr.magic != XFS_DIR2_LEAF1_MAGIC)
+               return __this_address;
+
+       return xfs_dir3_leaf_check_int(dp->i_mount, dp, &leafhdr, leaf);
+}
+
+static inline void
+xfs_dir3_leaf_check(
+       struct xfs_inode        *dp,
+       struct xfs_buf          *bp)
+{
+       xfs_failaddr_t          fa;
+
+       fa = xfs_dir3_leaf1_check(dp, bp);
+       if (!fa)
+               return;
+       xfs_corruption_error(__func__, XFS_ERRLEVEL_LOW, dp->i_mount,
+                       bp->b_addr, BBTOB(bp->b_length), __FILE__, __LINE__,
+                       fa);
+       ASSERT(0);
+}
+#else
+#define        xfs_dir3_leaf_check(dp, bp)
+#endif
+
+xfs_failaddr_t
+xfs_dir3_leaf_check_int(
+       struct xfs_mount        *mp,
+       struct xfs_inode        *dp,
+       struct xfs_dir3_icleaf_hdr *hdr,
+       struct xfs_dir2_leaf    *leaf)
+{
+       struct xfs_dir2_leaf_entry *ents;
+       xfs_dir2_leaf_tail_t    *ltp;
+       int                     stale;
+       int                     i;
+       const struct xfs_dir_ops *ops;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+       struct xfs_da_geometry  *geo = mp->m_dir_geo;
+
+       /*
+        * we can be passed a null dp here from a verifier, so we need to go the
+        * hard way to get them.
+        */
+       ops = xfs_dir_get_ops(mp, dp);
+
+       if (!hdr) {
+               ops->leaf_hdr_from_disk(&leafhdr, leaf);
+               hdr = &leafhdr;
+       }
+
+       ents = ops->leaf_ents_p(leaf);
+       ltp = xfs_dir2_leaf_tail_p(geo, leaf);
 
-static void xfs_dir2_leaf_log_bests(xfs_trans_t *, xfs_dabuf_t *, int, int);
-static void xfs_dir2_leaf_log_tail(xfs_trans_t *, xfs_dabuf_t *);
+       /*
+        * XXX (dgc): This value is not restrictive enough.
+        * Should factor in the size of the bests table as well.
+        * We can deduce a value for that from di_size.
+        */
+       if (hdr->count > ops->leaf_max_ents(geo))
+               return __this_address;
+
+       /* Leaves and bests don't overlap in leaf format. */
+       if ((hdr->magic == XFS_DIR2_LEAF1_MAGIC ||
+            hdr->magic == XFS_DIR3_LEAF1_MAGIC) &&
+           (char *)&ents[hdr->count] > (char *)xfs_dir2_leaf_bests_p(ltp))
+               return __this_address;
+
+       /* Check hash value order, count stale entries.  */
+       for (i = stale = 0; i < hdr->count; i++) {
+               if (i + 1 < hdr->count) {
+                       if (be32_to_cpu(ents[i].hashval) >
+                                       be32_to_cpu(ents[i + 1].hashval))
+                               return __this_address;
+               }
+               if (ents[i].address == cpu_to_be32(XFS_DIR2_NULL_DATAPTR))
+                       stale++;
+       }
+       if (hdr->stale != stale)
+               return __this_address;
+       return NULL;
+}
+
+/*
+ * We verify the magic numbers before decoding the leaf header so that on debug
+ * kernels we don't get assertion failures in xfs_dir3_leaf_hdr_from_disk() due
+ * to incorrect magic numbers.
+ */
+static xfs_failaddr_t
+xfs_dir3_leaf_verify(
+       struct xfs_buf          *bp,
+       uint16_t                magic)
+{
+       struct xfs_mount        *mp = bp->b_target->bt_mount;
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
+
+       ASSERT(magic == XFS_DIR2_LEAF1_MAGIC || magic == XFS_DIR2_LEAFN_MAGIC);
+
+       if (xfs_sb_version_hascrc(&mp->m_sb)) {
+               struct xfs_dir3_leaf_hdr *leaf3 = bp->b_addr;
+               uint16_t                magic3;
+
+               magic3 = (magic == XFS_DIR2_LEAF1_MAGIC) ? XFS_DIR3_LEAF1_MAGIC
+                                                        : XFS_DIR3_LEAFN_MAGIC;
+
+               if (leaf3->info.hdr.magic != cpu_to_be16(magic3))
+                       return __this_address;
+               if (!uuid_equal(&leaf3->info.uuid, &mp->m_sb.sb_meta_uuid))
+                       return __this_address;
+               if (be64_to_cpu(leaf3->info.blkno) != bp->b_bn)
+                       return __this_address;
+               if (!xfs_log_check_lsn(mp, be64_to_cpu(leaf3->info.lsn)))
+                       return __this_address;
+       } else {
+               if (leaf->hdr.info.magic != cpu_to_be16(magic))
+                       return __this_address;
+       }
+
+       return xfs_dir3_leaf_check_int(mp, NULL, NULL, leaf);
+}
+
+static void
+__read_verify(
+       struct xfs_buf  *bp,
+       uint16_t        magic)
+{
+       struct xfs_mount        *mp = bp->b_target->bt_mount;
+       xfs_failaddr_t          fa;
+
+       if (xfs_sb_version_hascrc(&mp->m_sb) &&
+            !xfs_buf_verify_cksum(bp, XFS_DIR3_LEAF_CRC_OFF))
+               xfs_verifier_error(bp, -EFSBADCRC, __this_address);
+       else {
+               fa = xfs_dir3_leaf_verify(bp, magic);
+               if (fa)
+                       xfs_verifier_error(bp, -EFSCORRUPTED, fa);
+       }
+}
+
+static void
+__write_verify(
+       struct xfs_buf  *bp,
+       uint16_t        magic)
+{
+       struct xfs_mount        *mp = bp->b_target->bt_mount;
+       struct xfs_buf_log_item *bip = bp->b_log_item;
+       struct xfs_dir3_leaf_hdr *hdr3 = bp->b_addr;
+       xfs_failaddr_t          fa;
+
+       fa = xfs_dir3_leaf_verify(bp, magic);
+       if (fa) {
+               xfs_verifier_error(bp, -EFSCORRUPTED, fa);
+               return;
+       }
+
+       if (!xfs_sb_version_hascrc(&mp->m_sb))
+               return;
+
+       if (bip)
+               hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
+
+       xfs_buf_update_cksum(bp, XFS_DIR3_LEAF_CRC_OFF);
+}
+
+static xfs_failaddr_t
+xfs_dir3_leaf1_verify(
+       struct xfs_buf  *bp)
+{
+       return xfs_dir3_leaf_verify(bp, XFS_DIR2_LEAF1_MAGIC);
+}
+
+static void
+xfs_dir3_leaf1_read_verify(
+       struct xfs_buf  *bp)
+{
+       __read_verify(bp, XFS_DIR2_LEAF1_MAGIC);
+}
+
+static void
+xfs_dir3_leaf1_write_verify(
+       struct xfs_buf  *bp)
+{
+       __write_verify(bp, XFS_DIR2_LEAF1_MAGIC);
+}
+
+static xfs_failaddr_t
+xfs_dir3_leafn_verify(
+       struct xfs_buf  *bp)
+{
+       return xfs_dir3_leaf_verify(bp, XFS_DIR2_LEAFN_MAGIC);
+}
+
+static void
+xfs_dir3_leafn_read_verify(
+       struct xfs_buf  *bp)
+{
+       __read_verify(bp, XFS_DIR2_LEAFN_MAGIC);
+}
+
+static void
+xfs_dir3_leafn_write_verify(
+       struct xfs_buf  *bp)
+{
+       __write_verify(bp, XFS_DIR2_LEAFN_MAGIC);
+}
+
+const struct xfs_buf_ops xfs_dir3_leaf1_buf_ops = {
+       .name = "xfs_dir3_leaf1",
+       .verify_read = xfs_dir3_leaf1_read_verify,
+       .verify_write = xfs_dir3_leaf1_write_verify,
+       .verify_struct = xfs_dir3_leaf1_verify,
+};
+
+const struct xfs_buf_ops xfs_dir3_leafn_buf_ops = {
+       .name = "xfs_dir3_leafn",
+       .verify_read = xfs_dir3_leafn_read_verify,
+       .verify_write = xfs_dir3_leafn_write_verify,
+       .verify_struct = xfs_dir3_leafn_verify,
+};
+
+int
+xfs_dir3_leaf_read(
+       struct xfs_trans        *tp,
+       struct xfs_inode        *dp,
+       xfs_dablk_t             fbno,
+       xfs_daddr_t             mappedbno,
+       struct xfs_buf          **bpp)
+{
+       int                     err;
+
+       err = xfs_da_read_buf(tp, dp, fbno, mappedbno, bpp,
+                               XFS_DATA_FORK, &xfs_dir3_leaf1_buf_ops);
+       if (!err && tp && *bpp)
+               xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_DIR_LEAF1_BUF);
+       return err;
+}
+
+int
+xfs_dir3_leafn_read(
+       struct xfs_trans        *tp,
+       struct xfs_inode        *dp,
+       xfs_dablk_t             fbno,
+       xfs_daddr_t             mappedbno,
+       struct xfs_buf          **bpp)
+{
+       int                     err;
+
+       err = xfs_da_read_buf(tp, dp, fbno, mappedbno, bpp,
+                               XFS_DATA_FORK, &xfs_dir3_leafn_buf_ops);
+       if (!err && tp && *bpp)
+               xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_DIR_LEAFN_BUF);
+       return err;
+}
+
+/*
+ * Initialize a new leaf block, leaf1 or leafn magic accepted.
+ */
+static void
+xfs_dir3_leaf_init(
+       struct xfs_mount        *mp,
+       struct xfs_trans        *tp,
+       struct xfs_buf          *bp,
+       xfs_ino_t               owner,
+       uint16_t                type)
+{
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
+
+       ASSERT(type == XFS_DIR2_LEAF1_MAGIC || type == XFS_DIR2_LEAFN_MAGIC);
+
+       if (xfs_sb_version_hascrc(&mp->m_sb)) {
+               struct xfs_dir3_leaf_hdr *leaf3 = bp->b_addr;
+
+               memset(leaf3, 0, sizeof(*leaf3));
+
+               leaf3->info.hdr.magic = (type == XFS_DIR2_LEAF1_MAGIC)
+                                        ? cpu_to_be16(XFS_DIR3_LEAF1_MAGIC)
+                                        : cpu_to_be16(XFS_DIR3_LEAFN_MAGIC);
+               leaf3->info.blkno = cpu_to_be64(bp->b_bn);
+               leaf3->info.owner = cpu_to_be64(owner);
+               uuid_copy(&leaf3->info.uuid, &mp->m_sb.sb_meta_uuid);
+       } else {
+               memset(leaf, 0, sizeof(*leaf));
+               leaf->hdr.info.magic = cpu_to_be16(type);
+       }
+
+       /*
+        * If it's a leaf-format directory initialize the tail.
+        * Caller is responsible for initialising the bests table.
+        */
+       if (type == XFS_DIR2_LEAF1_MAGIC) {
+               struct xfs_dir2_leaf_tail *ltp;
+
+               ltp = xfs_dir2_leaf_tail_p(mp->m_dir_geo, leaf);
+               ltp->bestcount = 0;
+               bp->b_ops = &xfs_dir3_leaf1_buf_ops;
+               xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAF1_BUF);
+       } else {
+               bp->b_ops = &xfs_dir3_leafn_buf_ops;
+               xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
+       }
+}
+
+int
+xfs_dir3_leaf_get_buf(
+       xfs_da_args_t           *args,
+       xfs_dir2_db_t           bno,
+       struct xfs_buf          **bpp,
+       uint16_t                magic)
+{
+       struct xfs_inode        *dp = args->dp;
+       struct xfs_trans        *tp = args->trans;
+       struct xfs_mount        *mp = dp->i_mount;
+       struct xfs_buf          *bp;
+       int                     error;
+
+       ASSERT(magic == XFS_DIR2_LEAF1_MAGIC || magic == XFS_DIR2_LEAFN_MAGIC);
+       ASSERT(bno >= xfs_dir2_byte_to_db(args->geo, XFS_DIR2_LEAF_OFFSET) &&
+              bno < xfs_dir2_byte_to_db(args->geo, XFS_DIR2_FREE_OFFSET));
+
+       error = xfs_da_get_buf(tp, dp, xfs_dir2_db_to_da(args->geo, bno),
+                              -1, &bp, XFS_DATA_FORK);
+       if (error)
+               return error;
+
+       xfs_dir3_leaf_init(mp, tp, bp, dp->i_ino, magic);
+       xfs_dir3_leaf_log_header(args, bp);
+       if (magic == XFS_DIR2_LEAF1_MAGIC)
+               xfs_dir3_leaf_log_tail(args, bp);
+       *bpp = bp;
+       return 0;
+}
 
 /*
  * Convert a block form directory to a leaf form directory.
@@ -49,27 +387,29 @@ static void xfs_dir2_leaf_log_tail(xfs_trans_t *, xfs_dabuf_t *);
 int                                            /* error */
 xfs_dir2_block_to_leaf(
        xfs_da_args_t           *args,          /* operation arguments */
-       xfs_dabuf_t             *dbp)           /* input block's buffer */
+       struct xfs_buf          *dbp)           /* input block's buffer */
 {
-       xfs_dir2_data_off_t     *bestsp;        /* leaf's bestsp entries */
+       __be16                  *bestsp;        /* leaf's bestsp entries */
        xfs_dablk_t             blkno;          /* leaf block's bno */
-       xfs_dir2_block_t        *block;         /* block structure */
+       xfs_dir2_data_hdr_t     *hdr;           /* block header */
        xfs_dir2_leaf_entry_t   *blp;           /* block's leaf entries */
        xfs_dir2_block_tail_t   *btp;           /* block's tail */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
-       xfs_dabuf_t             *lbp;           /* leaf block's buffer */
+       struct xfs_buf          *lbp;           /* leaf block's buffer */
        xfs_dir2_db_t           ldb;            /* leaf block's bno */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf's tail */
-       xfs_mount_t             *mp;            /* filesystem mount point */
        int                     needlog;        /* need to log block header */
        int                     needscan;       /* need to rescan bestfree */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir2_data_free *bf;
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+
+       trace_xfs_dir2_block_to_leaf(args);
 
-       xfs_dir2_trace_args_b("block_to_leaf", args, dbp);
        dp = args->dp;
-       mp = dp->i_mount;
        tp = args->trans;
        /*
         * Add the leaf block to the inode.
@@ -79,68 +419,200 @@ xfs_dir2_block_to_leaf(
        if ((error = xfs_da_grow_inode(args, &blkno))) {
                return error;
        }
-       ldb = XFS_DIR2_DA_TO_DB(mp, blkno);
-       ASSERT(ldb == XFS_DIR2_LEAF_FIRSTDB(mp));
+       ldb = xfs_dir2_da_to_db(args->geo, blkno);
+       ASSERT(ldb == xfs_dir2_byte_to_db(args->geo, XFS_DIR2_LEAF_OFFSET));
        /*
         * Initialize the leaf block, get a buffer for it.
         */
-       if ((error = xfs_dir2_leaf_init(args, ldb, &lbp, XFS_DIR2_LEAF1_MAGIC))) {
+       error = xfs_dir3_leaf_get_buf(args, ldb, &lbp, XFS_DIR2_LEAF1_MAGIC);
+       if (error)
                return error;
-       }
-       ASSERT(lbp != NULL);
-       leaf = lbp->data;
-       block = dbp->data;
-       xfs_dir2_data_check(dp, dbp);
-       btp = XFS_DIR2_BLOCK_TAIL_P(mp, block);
-       blp = XFS_DIR2_BLOCK_LEAF_P(btp);
+
+       leaf = lbp->b_addr;
+       hdr = dbp->b_addr;
+       xfs_dir3_data_check(dp, dbp);
+       btp = xfs_dir2_block_tail_p(args->geo, hdr);
+       blp = xfs_dir2_block_leaf_p(btp);
+       bf = dp->d_ops->data_bestfree_p(hdr);
+       ents = dp->d_ops->leaf_ents_p(leaf);
+
        /*
         * Set the counts in the leaf header.
         */
-       INT_COPY(leaf->hdr.count, btp->count, ARCH_CONVERT); /* INT_: type change */
-       INT_COPY(leaf->hdr.stale, btp->stale, ARCH_CONVERT); /* INT_: type change */
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+       leafhdr.count = be32_to_cpu(btp->count);
+       leafhdr.stale = be32_to_cpu(btp->stale);
+       dp->d_ops->leaf_hdr_to_disk(leaf, &leafhdr);
+       xfs_dir3_leaf_log_header(args, lbp);
+
        /*
         * Could compact these but I think we always do the conversion
         * after squeezing out stale entries.
         */
-       memcpy(leaf->ents, blp, INT_GET(btp->count, ARCH_CONVERT) * sizeof(xfs_dir2_leaf_entry_t));
-       xfs_dir2_leaf_log_ents(tp, lbp, 0, INT_GET(leaf->hdr.count, ARCH_CONVERT) - 1);
+       memcpy(ents, blp, be32_to_cpu(btp->count) * sizeof(xfs_dir2_leaf_entry_t));
+       xfs_dir3_leaf_log_ents(args, lbp, 0, leafhdr.count - 1);
        needscan = 0;
        needlog = 1;
        /*
         * Make the space formerly occupied by the leaf entries and block
         * tail be free.
         */
-       xfs_dir2_data_make_free(tp, dbp,
-               (xfs_dir2_data_aoff_t)((char *)blp - (char *)block),
-               (xfs_dir2_data_aoff_t)((char *)block + mp->m_dirblksize -
+       xfs_dir2_data_make_free(args, dbp,
+               (xfs_dir2_data_aoff_t)((char *)blp - (char *)hdr),
+               (xfs_dir2_data_aoff_t)((char *)hdr + args->geo->blksize -
                                       (char *)blp),
                &needlog, &needscan);
        /*
         * Fix up the block header, make it a data block.
         */
-       INT_SET(block->hdr.magic, ARCH_CONVERT, XFS_DIR2_DATA_MAGIC);
+       dbp->b_ops = &xfs_dir3_data_buf_ops;
+       xfs_trans_buf_set_type(tp, dbp, XFS_BLFT_DIR_DATA_BUF);
+       if (hdr->magic == cpu_to_be32(XFS_DIR2_BLOCK_MAGIC))
+               hdr->magic = cpu_to_be32(XFS_DIR2_DATA_MAGIC);
+       else
+               hdr->magic = cpu_to_be32(XFS_DIR3_DATA_MAGIC);
+
        if (needscan)
-               xfs_dir2_data_freescan(mp, (xfs_dir2_data_t *)block, &needlog,
-                       NULL);
+               xfs_dir2_data_freescan(dp, hdr, &needlog);
        /*
         * Set up leaf tail and bests table.
         */
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       INT_SET(ltp->bestcount, ARCH_CONVERT, 1);
-       bestsp = XFS_DIR2_LEAF_BESTS_P(ltp);
-       INT_COPY(bestsp[0], block->hdr.bestfree[0].length, ARCH_CONVERT);
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       ltp->bestcount = cpu_to_be32(1);
+       bestsp = xfs_dir2_leaf_bests_p(ltp);
+       bestsp[0] =  bf[0].length;
        /*
         * Log the data header and leaf bests table.
         */
        if (needlog)
-               xfs_dir2_data_log_header(tp, dbp);
-       xfs_dir2_leaf_check(dp, lbp);
-       xfs_dir2_data_check(dp, dbp);
-       xfs_dir2_leaf_log_bests(tp, lbp, 0, 0);
-       xfs_da_buf_done(lbp);
+               xfs_dir2_data_log_header(args, dbp);
+       xfs_dir3_leaf_check(dp, lbp);
+       xfs_dir3_data_check(dp, dbp);
+       xfs_dir3_leaf_log_bests(args, lbp, 0, 0);
        return 0;
 }
 
+STATIC void
+xfs_dir3_leaf_find_stale(
+       struct xfs_dir3_icleaf_hdr *leafhdr,
+       struct xfs_dir2_leaf_entry *ents,
+       int                     index,
+       int                     *lowstale,
+       int                     *highstale)
+{
+       /*
+        * Find the first stale entry before our index, if any.
+        */
+       for (*lowstale = index - 1; *lowstale >= 0; --*lowstale) {
+               if (ents[*lowstale].address ==
+                   cpu_to_be32(XFS_DIR2_NULL_DATAPTR))
+                       break;
+       }
+
+       /*
+        * Find the first stale entry at or after our index, if any.
+        * Stop if the result would require moving more entries than using
+        * lowstale.
+        */
+       for (*highstale = index; *highstale < leafhdr->count; ++*highstale) {
+               if (ents[*highstale].address ==
+                   cpu_to_be32(XFS_DIR2_NULL_DATAPTR))
+                       break;
+               if (*lowstale >= 0 && index - *lowstale <= *highstale - index)
+                       break;
+       }
+}
+
+struct xfs_dir2_leaf_entry *
+xfs_dir3_leaf_find_entry(
+       struct xfs_dir3_icleaf_hdr *leafhdr,
+       struct xfs_dir2_leaf_entry *ents,
+       int                     index,          /* leaf table position */
+       int                     compact,        /* need to compact leaves */
+       int                     lowstale,       /* index of prev stale leaf */
+       int                     highstale,      /* index of next stale leaf */
+       int                     *lfloglow,      /* low leaf logging index */
+       int                     *lfloghigh)     /* high leaf logging index */
+{
+       if (!leafhdr->stale) {
+               xfs_dir2_leaf_entry_t   *lep;   /* leaf entry table pointer */
+
+               /*
+                * Now we need to make room to insert the leaf entry.
+                *
+                * If there are no stale entries, just insert a hole at index.
+                */
+               lep = &ents[index];
+               if (index < leafhdr->count)
+                       memmove(lep + 1, lep,
+                               (leafhdr->count - index) * sizeof(*lep));
+
+               /*
+                * Record low and high logging indices for the leaf.
+                */
+               *lfloglow = index;
+               *lfloghigh = leafhdr->count++;
+               return lep;
+       }
+
+       /*
+        * There are stale entries.
+        *
+        * We will use one of them for the new entry.  It's probably not at
+        * the right location, so we'll have to shift some up or down first.
+        *
+        * If we didn't compact before, we need to find the nearest stale
+        * entries before and after our insertion point.
+        */
+       if (compact == 0)
+               xfs_dir3_leaf_find_stale(leafhdr, ents, index,
+                                        &lowstale, &highstale);
+
+       /*
+        * If the low one is better, use it.
+        */
+       if (lowstale >= 0 &&
+           (highstale == leafhdr->count ||
+            index - lowstale - 1 < highstale - index)) {
+               ASSERT(index - lowstale - 1 >= 0);
+               ASSERT(ents[lowstale].address ==
+                      cpu_to_be32(XFS_DIR2_NULL_DATAPTR));
+
+               /*
+                * Copy entries up to cover the stale entry and make room
+                * for the new entry.
+                */
+               if (index - lowstale - 1 > 0) {
+                       memmove(&ents[lowstale], &ents[lowstale + 1],
+                               (index - lowstale - 1) *
+                                       sizeof(xfs_dir2_leaf_entry_t));
+               }
+               *lfloglow = min(lowstale, *lfloglow);
+               *lfloghigh = max(index - 1, *lfloghigh);
+               leafhdr->stale--;
+               return &ents[index - 1];
+       }
+
+       /*
+        * The high one is better, so use that one.
+        */
+       ASSERT(highstale - index >= 0);
+       ASSERT(ents[highstale].address == cpu_to_be32(XFS_DIR2_NULL_DATAPTR));
+
+       /*
+        * Copy entries down to cover the stale entry and make room for the
+        * new entry.
+        */
+       if (highstale - index > 0) {
+               memmove(&ents[index + 1], &ents[index],
+                       (highstale - index) * sizeof(xfs_dir2_leaf_entry_t));
+       }
+       *lfloglow = min(index, *lfloglow);
+       *lfloghigh = max(highstale, *lfloghigh);
+       leafhdr->stale--;
+       return &ents[index];
+}
+
 /*
  * Add an entry to a leaf form directory.
  */
@@ -148,10 +620,10 @@ int                                               /* error */
 xfs_dir2_leaf_addname(
        xfs_da_args_t           *args)          /* operation arguments */
 {
-       xfs_dir2_data_off_t     *bestsp;        /* freespace table in leaf */
+       __be16                  *bestsp;        /* freespace table in leaf */
        int                     compact;        /* need to compact leaves */
-       xfs_dir2_data_t         *data;          /* data block structure */
-       xfs_dabuf_t             *dbp;           /* data block buffer */
+       xfs_dir2_data_hdr_t     *hdr;           /* data block header */
+       struct xfs_buf          *dbp;           /* data block buffer */
        xfs_dir2_data_entry_t   *dep;           /* data block entry */
        xfs_inode_t             *dp;            /* incore directory inode */
        xfs_dir2_data_unused_t  *dup;           /* data unused entry */
@@ -160,7 +632,7 @@ xfs_dir2_leaf_addname(
        int                     highstale;      /* index of next stale leaf */
        int                     i;              /* temporary, index */
        int                     index;          /* leaf table position */
-       xfs_dabuf_t             *lbp;           /* leaf's buffer */
+       struct xfs_buf          *lbp;           /* leaf's buffer */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        int                     length;         /* length of new entry */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry table pointer */
@@ -168,27 +640,25 @@ xfs_dir2_leaf_addname(
        int                     lfloghigh;      /* high leaf logging index */
        int                     lowstale;       /* index of prev stale leaf */
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail pointer */
-       xfs_mount_t             *mp;            /* filesystem mount point */
        int                     needbytes;      /* leaf block bytes needed */
        int                     needlog;        /* need to log data header */
        int                     needscan;       /* need to rescan data free */
-       xfs_dir2_data_off_t     *tagp;          /* end of data entry */
+       __be16                  *tagp;          /* end of data entry */
        xfs_trans_t             *tp;            /* transaction pointer */
        xfs_dir2_db_t           use_block;      /* data block number */
+       struct xfs_dir2_data_free *bf;          /* bestfree table */
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+
+       trace_xfs_dir2_leaf_addname(args);
 
-       xfs_dir2_trace_args("leaf_addname", args);
        dp = args->dp;
        tp = args->trans;
-       mp = dp->i_mount;
-       /*
-        * Read the leaf block.
-        */
-       error = xfs_da_read_buf(tp, dp, mp->m_dirleafblk, -1, &lbp,
-               XFS_DATA_FORK);
-       if (error) {
+
+       error = xfs_dir3_leaf_read(tp, dp, args->geo->leafblk, -1, &lbp);
+       if (error)
                return error;
-       }
-       ASSERT(lbp != NULL);
+
        /*
         * Look up the entry by hash value and name.
         * We know it's not there, our caller has already done a lookup.
@@ -196,25 +666,28 @@ xfs_dir2_leaf_addname(
         * But if there are dup hash values the index is of the first of those.
         */
        index = xfs_dir2_leaf_search_hash(args, lbp);
-       leaf = lbp->data;
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       bestsp = XFS_DIR2_LEAF_BESTS_P(ltp);
-       length = XFS_DIR2_DATA_ENTSIZE(args->namelen);
+       leaf = lbp->b_addr;
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       ents = dp->d_ops->leaf_ents_p(leaf);
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+       bestsp = xfs_dir2_leaf_bests_p(ltp);
+       length = dp->d_ops->data_entsize(args->namelen);
+
        /*
         * See if there are any entries with the same hash value
         * and space in their block for the new entry.
         * This is good because it puts multiple same-hash value entries
         * in a data block, improving the lookup of those entries.
         */
-       for (use_block = -1, lep = &leaf->ents[index];
-            index < INT_GET(leaf->hdr.count, ARCH_CONVERT) && INT_GET(lep->hashval, ARCH_CONVERT) == args->hashval;
+       for (use_block = -1, lep = &ents[index];
+            index < leafhdr.count && be32_to_cpu(lep->hashval) == args->hashval;
             index++, lep++) {
-               if (INT_GET(lep->address, ARCH_CONVERT) == XFS_DIR2_NULL_DATAPTR)
+               if (be32_to_cpu(lep->address) == XFS_DIR2_NULL_DATAPTR)
                        continue;
-               i = XFS_DIR2_DATAPTR_TO_DB(mp, INT_GET(lep->address, ARCH_CONVERT));
-               ASSERT(i < INT_GET(ltp->bestcount, ARCH_CONVERT));
-               ASSERT(INT_GET(bestsp[i], ARCH_CONVERT) != NULLDATAOFF);
-               if (INT_GET(bestsp[i], ARCH_CONVERT) >= length) {
+               i = xfs_dir2_dataptr_to_db(args->geo, be32_to_cpu(lep->address));
+               ASSERT(i < be32_to_cpu(ltp->bestcount));
+               ASSERT(bestsp[i] != cpu_to_be16(NULLDATAOFF));
+               if (be16_to_cpu(bestsp[i]) >= length) {
                        use_block = i;
                        break;
                }
@@ -223,13 +696,14 @@ xfs_dir2_leaf_addname(
         * Didn't find a block yet, linear search all the data blocks.
         */
        if (use_block == -1) {
-               for (i = 0; i < INT_GET(ltp->bestcount, ARCH_CONVERT); i++) {
+               for (i = 0; i < be32_to_cpu(ltp->bestcount); i++) {
                        /*
                         * Remember a block we see that's missing.
                         */
-                       if (INT_GET(bestsp[i], ARCH_CONVERT) == NULLDATAOFF && use_block == -1)
+                       if (bestsp[i] == cpu_to_be16(NULLDATAOFF) &&
+                           use_block == -1)
                                use_block = i;
-                       else if (INT_GET(bestsp[i], ARCH_CONVERT) >= length) {
+                       else if (be16_to_cpu(bestsp[i]) >= length) {
                                use_block = i;
                                break;
                        }
@@ -238,41 +712,43 @@ xfs_dir2_leaf_addname(
        /*
         * How many bytes do we need in the leaf block?
         */
-       needbytes =
-               (leaf->hdr.stale ? 0 : (uint)sizeof(leaf->ents[0])) +
-               (use_block != -1 ? 0 : (uint)sizeof(leaf->bests[0]));
+       needbytes = 0;
+       if (!leafhdr.stale)
+               needbytes += sizeof(xfs_dir2_leaf_entry_t);
+       if (use_block == -1)
+               needbytes += sizeof(xfs_dir2_data_off_t);
+
        /*
         * Now kill use_block if it refers to a missing block, so we
         * can use it as an indication of allocation needed.
         */
-       if (use_block != -1 && INT_GET(bestsp[use_block], ARCH_CONVERT) == NULLDATAOFF)
+       if (use_block != -1 && bestsp[use_block] == cpu_to_be16(NULLDATAOFF))
                use_block = -1;
        /*
         * If we don't have enough free bytes but we can make enough
         * by compacting out stale entries, we'll do that.
         */
-       if ((char *)bestsp - (char *)&leaf->ents[INT_GET(leaf->hdr.count, ARCH_CONVERT)] < needbytes &&
-           INT_GET(leaf->hdr.stale, ARCH_CONVERT) > 1) {
+       if ((char *)bestsp - (char *)&ents[leafhdr.count] < needbytes &&
+           leafhdr.stale > 1)
                compact = 1;
-       }
+
        /*
         * Otherwise if we don't have enough free bytes we need to
         * convert to node form.
         */
-       else if ((char *)bestsp - (char *)&leaf->ents[INT_GET(leaf->hdr.count, ARCH_CONVERT)] <
-                needbytes) {
+       else if ((char *)bestsp - (char *)&ents[leafhdr.count] < needbytes) {
                /*
                 * Just checking or no space reservation, give up.
                 */
-               if (args->justcheck || args->total == 0) {
-                       xfs_da_brelse(tp, lbp);
-                       return XFS_ERROR(ENOSPC);
+               if ((args->op_flags & XFS_DA_OP_JUSTCHECK) ||
+                                                       args->total == 0) {
+                       xfs_trans_brelse(tp, lbp);
+                       return -ENOSPC;
                }
                /*
                 * Convert to node form.
                 */
                error = xfs_dir2_leaf_to_node(args, lbp);
-               xfs_da_buf_done(lbp);
                if (error)
                        return error;
                /*
@@ -289,17 +765,17 @@ xfs_dir2_leaf_addname(
         * If just checking, then it will fit unless we needed to allocate
         * a new data block.
         */
-       if (args->justcheck) {
-               xfs_da_brelse(tp, lbp);
-               return use_block == -1 ? XFS_ERROR(ENOSPC) : 0;
+       if (args->op_flags & XFS_DA_OP_JUSTCHECK) {
+               xfs_trans_brelse(tp, lbp);
+               return use_block == -1 ? -ENOSPC : 0;
        }
        /*
         * If no allocations are allowed, return now before we've
         * changed anything.
         */
        if (args->total == 0 && use_block == -1) {
-               xfs_da_brelse(tp, lbp);
-               return XFS_ERROR(ENOSPC);
+               xfs_trans_brelse(tp, lbp);
+               return -ENOSPC;
        }
        /*
         * Need to compact the leaf entries, removing stale ones.
@@ -308,15 +784,15 @@ xfs_dir2_leaf_addname(
         * point later.
         */
        if (compact) {
-               xfs_dir2_leaf_compact_x1(lbp, &index, &lowstale, &highstale,
-                       &lfloglow, &lfloghigh);
+               xfs_dir3_leaf_compact_x1(&leafhdr, ents, &index, &lowstale,
+                       &highstale, &lfloglow, &lfloghigh);
        }
        /*
         * There are stale entries, so we'll need log-low and log-high
         * impossibly bad values later.
         */
-       else if (INT_GET(leaf->hdr.stale, ARCH_CONVERT)) {
-               lfloglow = INT_GET(leaf->hdr.count, ARCH_CONVERT);
+       else if (leafhdr.stale) {
+               lfloglow = leafhdr.count;
                lfloghigh = -1;
        }
        /*
@@ -329,275 +805,149 @@ xfs_dir2_leaf_addname(
                 */
                if ((error = xfs_dir2_grow_inode(args, XFS_DIR2_DATA_SPACE,
                                &use_block))) {
-                       xfs_da_brelse(tp, lbp);
+                       xfs_trans_brelse(tp, lbp);
                        return error;
                }
                /*
                 * Initialize the block.
                 */
-               if ((error = xfs_dir2_data_init(args, use_block, &dbp))) {
-                       xfs_da_brelse(tp, lbp);
+               if ((error = xfs_dir3_data_init(args, use_block, &dbp))) {
+                       xfs_trans_brelse(tp, lbp);
                        return error;
                }
                /*
                 * If we're adding a new data block on the end we need to
                 * extend the bests table.  Copy it up one entry.
                 */
-               if (use_block >= INT_GET(ltp->bestcount, ARCH_CONVERT)) {
+               if (use_block >= be32_to_cpu(ltp->bestcount)) {
                        bestsp--;
                        memmove(&bestsp[0], &bestsp[1],
-                               INT_GET(ltp->bestcount, ARCH_CONVERT) * sizeof(bestsp[0]));
-                       INT_MOD(ltp->bestcount, ARCH_CONVERT, +1);
-                       xfs_dir2_leaf_log_tail(tp, lbp);
-                       xfs_dir2_leaf_log_bests(tp, lbp, 0, INT_GET(ltp->bestcount, ARCH_CONVERT) - 1);
+                               be32_to_cpu(ltp->bestcount) * sizeof(bestsp[0]));
+                       be32_add_cpu(&ltp->bestcount, 1);
+                       xfs_dir3_leaf_log_tail(args, lbp);
+                       xfs_dir3_leaf_log_bests(args, lbp, 0,
+                                               be32_to_cpu(ltp->bestcount) - 1);
                }
                /*
                 * If we're filling in a previously empty block just log it.
                 */
                else
-                       xfs_dir2_leaf_log_bests(tp, lbp, use_block, use_block);
-               data = dbp->data;
-               INT_COPY(bestsp[use_block], data->hdr.bestfree[0].length, ARCH_CONVERT);
+                       xfs_dir3_leaf_log_bests(args, lbp, use_block, use_block);
+               hdr = dbp->b_addr;
+               bf = dp->d_ops->data_bestfree_p(hdr);
+               bestsp[use_block] = bf[0].length;
                grown = 1;
-       }
-       /*
-        * Already had space in some data block.
-        * Just read that one in.
-        */
-       else {
-               if ((error =
-                   xfs_da_read_buf(tp, dp, XFS_DIR2_DB_TO_DA(mp, use_block),
-                           -1, &dbp, XFS_DATA_FORK))) {
-                       xfs_da_brelse(tp, lbp);
+       } else {
+               /*
+                * Already had space in some data block.
+                * Just read that one in.
+                */
+               error = xfs_dir3_data_read(tp, dp,
+                                  xfs_dir2_db_to_da(args->geo, use_block),
+                                  -1, &dbp);
+               if (error) {
+                       xfs_trans_brelse(tp, lbp);
                        return error;
                }
-               data = dbp->data;
+               hdr = dbp->b_addr;
+               bf = dp->d_ops->data_bestfree_p(hdr);
                grown = 0;
        }
-       xfs_dir2_data_check(dp, dbp);
        /*
         * Point to the biggest freespace in our data block.
         */
        dup = (xfs_dir2_data_unused_t *)
-             ((char *)data + INT_GET(data->hdr.bestfree[0].offset, ARCH_CONVERT));
-       ASSERT(INT_GET(dup->length, ARCH_CONVERT) >= length);
+             ((char *)hdr + be16_to_cpu(bf[0].offset));
        needscan = needlog = 0;
        /*
         * Mark the initial part of our freespace in use for the new entry.
         */
-       xfs_dir2_data_use_free(tp, dbp, dup,
-               (xfs_dir2_data_aoff_t)((char *)dup - (char *)data), length,
-               &needlog, &needscan);
+       error = xfs_dir2_data_use_free(args, dbp, dup,
+                       (xfs_dir2_data_aoff_t)((char *)dup - (char *)hdr),
+                       length, &needlog, &needscan);
+       if (error) {
+               xfs_trans_brelse(tp, lbp);
+               return error;
+       }
        /*
         * Initialize our new entry (at last).
         */
        dep = (xfs_dir2_data_entry_t *)dup;
-       INT_SET(dep->inumber, ARCH_CONVERT, args->inumber);
+       dep->inumber = cpu_to_be64(args->inumber);
        dep->namelen = args->namelen;
        memcpy(dep->name, args->name, dep->namelen);
-       tagp = XFS_DIR2_DATA_ENTRY_TAG_P(dep);
-       INT_SET(*tagp, ARCH_CONVERT, (xfs_dir2_data_off_t)((char *)dep - (char *)data));
+       dp->d_ops->data_put_ftype(dep, args->filetype);
+       tagp = dp->d_ops->data_entry_tag_p(dep);
+       *tagp = cpu_to_be16((char *)dep - (char *)hdr);
        /*
         * Need to scan fix up the bestfree table.
         */
        if (needscan)
-               xfs_dir2_data_freescan(mp, data, &needlog, NULL);
+               xfs_dir2_data_freescan(dp, hdr, &needlog);
        /*
         * Need to log the data block's header.
         */
        if (needlog)
-               xfs_dir2_data_log_header(tp, dbp);
-       xfs_dir2_data_log_entry(tp, dbp, dep);
+               xfs_dir2_data_log_header(args, dbp);
+       xfs_dir2_data_log_entry(args, dbp, dep);
        /*
         * If the bests table needs to be changed, do it.
         * Log the change unless we've already done that.
         */
-       if (INT_GET(bestsp[use_block], ARCH_CONVERT) != INT_GET(data->hdr.bestfree[0].length, ARCH_CONVERT)) {
-               INT_COPY(bestsp[use_block], data->hdr.bestfree[0].length, ARCH_CONVERT);
+       if (be16_to_cpu(bestsp[use_block]) != be16_to_cpu(bf[0].length)) {
+               bestsp[use_block] = bf[0].length;
                if (!grown)
-                       xfs_dir2_leaf_log_bests(tp, lbp, use_block, use_block);
-       }
-       /*
-        * Now we need to make room to insert the leaf entry.
-        * If there are no stale entries, we just insert a hole at index.
-        */
-       if (!leaf->hdr.stale) {
-               /*
-                * lep is still good as the index leaf entry.
-                */
-               if (index < INT_GET(leaf->hdr.count, ARCH_CONVERT))
-                       memmove(lep + 1, lep,
-                               (INT_GET(leaf->hdr.count, ARCH_CONVERT) - index) * sizeof(*lep));
-               /*
-                * Record low and high logging indices for the leaf.
-                */
-               lfloglow = index;
-               lfloghigh = INT_GET(leaf->hdr.count, ARCH_CONVERT);
-               INT_MOD(leaf->hdr.count, ARCH_CONVERT, +1);
-       }
-       /*
-        * There are stale entries.
-        * We will use one of them for the new entry.
-        * It's probably not at the right location, so we'll have to
-        * shift some up or down first.
-        */
-       else {
-               /*
-                * If we didn't compact before, we need to find the nearest
-                * stale entries before and after our insertion point.
-                */
-               if (compact == 0) {
-                       /*
-                        * Find the first stale entry before the insertion
-                        * point, if any.
-                        */
-                       for (lowstale = index - 1;
-                            lowstale >= 0 &&
-                               INT_GET(leaf->ents[lowstale].address, ARCH_CONVERT) !=
-                               XFS_DIR2_NULL_DATAPTR;
-                            lowstale--)
-                               continue;
-                       /*
-                        * Find the next stale entry at or after the insertion
-                        * point, if any.   Stop if we go so far that the
-                        * lowstale entry would be better.
-                        */
-                       for (highstale = index;
-                            highstale < INT_GET(leaf->hdr.count, ARCH_CONVERT) &&
-                               INT_GET(leaf->ents[highstale].address, ARCH_CONVERT) !=
-                               XFS_DIR2_NULL_DATAPTR &&
-                               (lowstale < 0 ||
-                                index - lowstale - 1 >= highstale - index);
-                            highstale++)
-                               continue;
-               }
-               /*
-                * If the low one is better, use it.
-                */
-               if (lowstale >= 0 &&
-                   (highstale == INT_GET(leaf->hdr.count, ARCH_CONVERT) ||
-                    index - lowstale - 1 < highstale - index)) {
-                       ASSERT(index - lowstale - 1 >= 0);
-                       ASSERT(INT_GET(leaf->ents[lowstale].address, ARCH_CONVERT) ==
-                              XFS_DIR2_NULL_DATAPTR);
-                       /*
-                        * Copy entries up to cover the stale entry
-                        * and make room for the new entry.
-                        */
-                       if (index - lowstale - 1 > 0)
-                               memmove(&leaf->ents[lowstale],
-                                       &leaf->ents[lowstale + 1],
-                                       (index - lowstale - 1) * sizeof(*lep));
-                       lep = &leaf->ents[index - 1];
-                       lfloglow = MIN(lowstale, lfloglow);
-                       lfloghigh = MAX(index - 1, lfloghigh);
-               }
-               /*
-                * The high one is better, so use that one.
-                */
-               else {
-                       ASSERT(highstale - index >= 0);
-                       ASSERT(INT_GET(leaf->ents[highstale].address, ARCH_CONVERT) ==
-                              XFS_DIR2_NULL_DATAPTR);
-                       /*
-                        * Copy entries down to copver the stale entry
-                        * and make room for the new entry.
-                        */
-                       if (highstale - index > 0)
-                               memmove(&leaf->ents[index + 1],
-                                       &leaf->ents[index],
-                                       (highstale - index) * sizeof(*lep));
-                       lep = &leaf->ents[index];
-                       lfloglow = MIN(index, lfloglow);
-                       lfloghigh = MAX(highstale, lfloghigh);
-               }
-               INT_MOD(leaf->hdr.stale, ARCH_CONVERT, -1);
+                       xfs_dir3_leaf_log_bests(args, lbp, use_block, use_block);
        }
+
+       lep = xfs_dir3_leaf_find_entry(&leafhdr, ents, index, compact, lowstale,
+                                      highstale, &lfloglow, &lfloghigh);
+
        /*
         * Fill in the new leaf entry.
         */
-       INT_SET(lep->hashval, ARCH_CONVERT, args->hashval);
-       INT_SET(lep->address, ARCH_CONVERT, XFS_DIR2_DB_OFF_TO_DATAPTR(mp, use_block, INT_GET(*tagp, ARCH_CONVERT)));
+       lep->hashval = cpu_to_be32(args->hashval);
+       lep->address = cpu_to_be32(
+                               xfs_dir2_db_off_to_dataptr(args->geo, use_block,
+                               be16_to_cpu(*tagp)));
        /*
         * Log the leaf fields and give up the buffers.
         */
-       xfs_dir2_leaf_log_header(tp, lbp);
-       xfs_dir2_leaf_log_ents(tp, lbp, lfloglow, lfloghigh);
-       xfs_dir2_leaf_check(dp, lbp);
-       xfs_da_buf_done(lbp);
-       xfs_dir2_data_check(dp, dbp);
-       xfs_da_buf_done(dbp);
+       dp->d_ops->leaf_hdr_to_disk(leaf, &leafhdr);
+       xfs_dir3_leaf_log_header(args, lbp);
+       xfs_dir3_leaf_log_ents(args, lbp, lfloglow, lfloghigh);
+       xfs_dir3_leaf_check(dp, lbp);
+       xfs_dir3_data_check(dp, dbp);
        return 0;
 }
 
-#ifdef DEBUG
-/*
- * Check the internal consistency of a leaf1 block.
- * Pop an assert if something is wrong.
- */
-void
-xfs_dir2_leaf_check(
-       xfs_inode_t             *dp,            /* incore directory inode */
-       xfs_dabuf_t             *bp)            /* leaf's buffer */
-{
-       int                     i;              /* leaf index */
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
-       xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail pointer */
-       xfs_mount_t             *mp;            /* filesystem mount point */
-       int                     stale;          /* count of stale leaves */
-
-       leaf = bp->data;
-       mp = dp->i_mount;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAF1_MAGIC);
-       /*
-        * This value is not restrictive enough.
-        * Should factor in the size of the bests table as well.
-        * We can deduce a value for that from di_size.
-        */
-       ASSERT(INT_GET(leaf->hdr.count, ARCH_CONVERT) <= XFS_DIR2_MAX_LEAF_ENTS(mp));
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       /*
-        * Leaves and bests don't overlap.
-        */
-       ASSERT((char *)&leaf->ents[INT_GET(leaf->hdr.count, ARCH_CONVERT)] <=
-              (char *)XFS_DIR2_LEAF_BESTS_P(ltp));
-       /*
-        * Check hash value order, count stale entries.
-        */
-       for (i = stale = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT); i++) {
-               if (i + 1 < INT_GET(leaf->hdr.count, ARCH_CONVERT))
-                       ASSERT(INT_GET(leaf->ents[i].hashval, ARCH_CONVERT) <=
-                              INT_GET(leaf->ents[i + 1].hashval, ARCH_CONVERT));
-               if (INT_GET(leaf->ents[i].address, ARCH_CONVERT) == XFS_DIR2_NULL_DATAPTR)
-                       stale++;
-       }
-       ASSERT(INT_GET(leaf->hdr.stale, ARCH_CONVERT) == stale);
-}
-#endif /* DEBUG */
-
 /*
  * Compact out any stale entries in the leaf.
  * Log the header and changed leaf entries, if any.
  */
 void
-xfs_dir2_leaf_compact(
+xfs_dir3_leaf_compact(
        xfs_da_args_t   *args,          /* operation arguments */
-       xfs_dabuf_t     *bp)            /* leaf buffer */
+       struct xfs_dir3_icleaf_hdr *leafhdr,
+       struct xfs_buf  *bp)            /* leaf buffer */
 {
        int             from;           /* source leaf index */
        xfs_dir2_leaf_t *leaf;          /* leaf structure */
        int             loglow;         /* first leaf entry to log */
        int             to;             /* target leaf index */
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_inode *dp = args->dp;
 
-       leaf = bp->data;
-       if (!leaf->hdr.stale) {
+       leaf = bp->b_addr;
+       if (!leafhdr->stale)
                return;
-       }
+
        /*
         * Compress out the stale entries in place.
         */
-       for (from = to = 0, loglow = -1; from < INT_GET(leaf->hdr.count, ARCH_CONVERT); from++) {
-               if (INT_GET(leaf->ents[from].address, ARCH_CONVERT) == XFS_DIR2_NULL_DATAPTR)
+       ents = dp->d_ops->leaf_ents_p(leaf);
+       for (from = to = 0, loglow = -1; from < leafhdr->count; from++) {
+               if (ents[from].address == cpu_to_be32(XFS_DIR2_NULL_DATAPTR))
                        continue;
                /*
                 * Only actually copy the entries that are different.
@@ -605,19 +955,21 @@ xfs_dir2_leaf_compact(
                if (from > to) {
                        if (loglow == -1)
                                loglow = to;
-                       leaf->ents[to] = leaf->ents[from];
+                       ents[to] = ents[from];
                }
                to++;
        }
        /*
         * Update and log the header, log the leaf entries.
         */
-       ASSERT(INT_GET(leaf->hdr.stale, ARCH_CONVERT) == from - to);
-       INT_MOD(leaf->hdr.count, ARCH_CONVERT, -(INT_GET(leaf->hdr.stale, ARCH_CONVERT)));
-       leaf->hdr.stale = 0;
-       xfs_dir2_leaf_log_header(args->trans, bp);
+       ASSERT(leafhdr->stale == from - to);
+       leafhdr->count -= leafhdr->stale;
+       leafhdr->stale = 0;
+
+       dp->d_ops->leaf_hdr_to_disk(leaf, leafhdr);
+       xfs_dir3_leaf_log_header(args, bp);
        if (loglow != -1)
-               xfs_dir2_leaf_log_ents(args->trans, bp, loglow, to - 1);
+               xfs_dir3_leaf_log_ents(args, bp, loglow, to - 1);
 }
 
 /*
@@ -629,8 +981,9 @@ xfs_dir2_leaf_compact(
  * and leaf logging indices.
  */
 void
-xfs_dir2_leaf_compact_x1(
-       xfs_dabuf_t     *bp,            /* leaf buffer */
+xfs_dir3_leaf_compact_x1(
+       struct xfs_dir3_icleaf_hdr *leafhdr,
+       struct xfs_dir2_leaf_entry *ents,
        int             *indexp,        /* insertion index */
        int             *lowstalep,     /* out: stale entry before us */
        int             *highstalep,    /* out: stale entry after us */
@@ -641,37 +994,20 @@ xfs_dir2_leaf_compact_x1(
        int             highstale;      /* stale entry at/after index */
        int             index;          /* insertion index */
        int             keepstale;      /* source index of kept stale */
-       xfs_dir2_leaf_t *leaf;          /* leaf structure */
        int             lowstale;       /* stale entry before index */
        int             newindex=0;     /* new insertion index */
        int             to;             /* destination copy index */
 
-       leaf = bp->data;
-       ASSERT(INT_GET(leaf->hdr.stale, ARCH_CONVERT) > 1);
+       ASSERT(leafhdr->stale > 1);
        index = *indexp;
-       /*
-        * Find the first stale entry before our index, if any.
-        */
-       for (lowstale = index - 1;
-            lowstale >= 0 &&
-               INT_GET(leaf->ents[lowstale].address, ARCH_CONVERT) != XFS_DIR2_NULL_DATAPTR;
-            lowstale--)
-               continue;
-       /*
-        * Find the first stale entry at or after our index, if any.
-        * Stop if the answer would be worse than lowstale.
-        */
-       for (highstale = index;
-            highstale < INT_GET(leaf->hdr.count, ARCH_CONVERT) &&
-               INT_GET(leaf->ents[highstale].address, ARCH_CONVERT) != XFS_DIR2_NULL_DATAPTR &&
-               (lowstale < 0 || index - lowstale > highstale - index);
-            highstale++)
-               continue;
+
+       xfs_dir3_leaf_find_stale(leafhdr, ents, index, &lowstale, &highstale);
+
        /*
         * Pick the better of lowstale and highstale.
         */
        if (lowstale >= 0 &&
-           (highstale == INT_GET(leaf->hdr.count, ARCH_CONVERT) ||
+           (highstale == leafhdr->count ||
             index - lowstale <= highstale - index))
                keepstale = lowstale;
        else
@@ -680,14 +1016,14 @@ xfs_dir2_leaf_compact_x1(
         * Copy the entries in place, removing all the stale entries
         * except keepstale.
         */
-       for (from = to = 0; from < INT_GET(leaf->hdr.count, ARCH_CONVERT); from++) {
+       for (from = to = 0; from < leafhdr->count; from++) {
                /*
                 * Notice the new value of index.
                 */
                if (index == from)
                        newindex = to;
                if (from != keepstale &&
-                   INT_GET(leaf->ents[from].address, ARCH_CONVERT) == XFS_DIR2_NULL_DATAPTR) {
+                   ents[from].address == cpu_to_be32(XFS_DIR2_NULL_DATAPTR)) {
                        if (from == to)
                                *lowlogp = to;
                        continue;
@@ -701,7 +1037,7 @@ xfs_dir2_leaf_compact_x1(
                 * Copy only the entries that have moved.
                 */
                if (from > to)
-                       leaf->ents[to] = leaf->ents[from];
+                       ents[to] = ents[from];
                to++;
        }
        ASSERT(from > to);
@@ -715,8 +1051,8 @@ xfs_dir2_leaf_compact_x1(
        /*
         * Adjust the leaf header values.
         */
-       INT_MOD(leaf->hdr.count, ARCH_CONVERT, -(from - to));
-       INT_SET(leaf->hdr.stale, ARCH_CONVERT, 1);
+       leafhdr->count -= from - to;
+       leafhdr->stale = 1;
        /*
         * Remember the low/high stale value only in the "right"
         * direction.
@@ -724,90 +1060,35 @@ xfs_dir2_leaf_compact_x1(
        if (lowstale >= newindex)
                lowstale = -1;
        else
-               highstale = INT_GET(leaf->hdr.count, ARCH_CONVERT);
-       *highlogp = INT_GET(leaf->hdr.count, ARCH_CONVERT) - 1;
+               highstale = leafhdr->count;
+       *highlogp = leafhdr->count - 1;
        *lowstalep = lowstale;
        *highstalep = highstale;
 }
 
-/*
- * Initialize a new leaf block, leaf1 or leafn magic accepted.
- */
-int
-xfs_dir2_leaf_init(
-       xfs_da_args_t           *args,          /* operation arguments */
-       xfs_dir2_db_t           bno,            /* directory block number */
-       xfs_dabuf_t             **bpp,          /* out: leaf buffer */
-       int                     magic)          /* magic number for block */
-{
-       xfs_dabuf_t             *bp;            /* leaf buffer */
-       xfs_inode_t             *dp;            /* incore directory inode */
-       int                     error;          /* error return code */
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
-       xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail structure */
-       xfs_mount_t             *mp;            /* filesystem mount point */
-       xfs_trans_t             *tp;            /* transaction pointer */
-
-       dp = args->dp;
-       ASSERT(dp != NULL);
-       tp = args->trans;
-       mp = dp->i_mount;
-       ASSERT(bno >= XFS_DIR2_LEAF_FIRSTDB(mp) &&
-              bno < XFS_DIR2_FREE_FIRSTDB(mp));
-       /*
-        * Get the buffer for the block.
-        */
-       error = xfs_da_get_buf(tp, dp, XFS_DIR2_DB_TO_DA(mp, bno), -1, &bp,
-               XFS_DATA_FORK);
-       if (error) {
-               return error;
-       }
-       ASSERT(bp != NULL);
-       leaf = bp->data;
-       /*
-        * Initialize the header.
-        */
-       INT_SET(leaf->hdr.info.magic, ARCH_CONVERT, magic);
-       leaf->hdr.info.forw = 0;
-       leaf->hdr.info.back = 0;
-       leaf->hdr.count = 0;
-       leaf->hdr.stale = 0;
-       xfs_dir2_leaf_log_header(tp, bp);
-       /*
-        * If it's a leaf-format directory initialize the tail.
-        * In this case our caller has the real bests table to copy into
-        * the block.
-        */
-       if (magic == XFS_DIR2_LEAF1_MAGIC) {
-               ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-               ltp->bestcount = 0;
-               xfs_dir2_leaf_log_tail(tp, bp);
-       }
-       *bpp = bp;
-       return 0;
-}
-
 /*
  * Log the bests entries indicated from a leaf1 block.
  */
-void
-xfs_dir2_leaf_log_bests(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_dabuf_t             *bp,            /* leaf buffer */
+static void
+xfs_dir3_leaf_log_bests(
+       struct xfs_da_args      *args,
+       struct xfs_buf          *bp,            /* leaf buffer */
        int                     first,          /* first entry to log */
        int                     last)           /* last entry to log */
 {
-       xfs_dir2_data_off_t     *firstb;        /* pointer to first entry */
-       xfs_dir2_data_off_t     *lastb;         /* pointer to last entry */
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
+       __be16                  *firstb;        /* pointer to first entry */
+       __be16                  *lastb;         /* pointer to last entry */
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail structure */
 
-       leaf = bp->data;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAF1_MAGIC);
-       ltp = XFS_DIR2_LEAF_TAIL_P(tp->t_mountp, leaf);
-       firstb = XFS_DIR2_LEAF_BESTS_P(ltp) + first;
-       lastb = XFS_DIR2_LEAF_BESTS_P(ltp) + last;
-       xfs_da_log_buf(tp, bp, (uint)((char *)firstb - (char *)leaf),
+       ASSERT(leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAF1_MAGIC));
+
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       firstb = xfs_dir2_leaf_bests_p(ltp) + first;
+       lastb = xfs_dir2_leaf_bests_p(ltp) + last;
+       xfs_trans_log_buf(args->trans, bp,
+               (uint)((char *)firstb - (char *)leaf),
                (uint)((char *)lastb - (char *)leaf + sizeof(*lastb) - 1));
 }
 
@@ -815,22 +1096,27 @@ xfs_dir2_leaf_log_bests(
  * Log the leaf entries indicated from a leaf1 or leafn block.
  */
 void
-xfs_dir2_leaf_log_ents(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_dabuf_t             *bp,            /* leaf buffer */
-       int                     first,          /* first entry to log */
-       int                     last)           /* last entry to log */
+xfs_dir3_leaf_log_ents(
+       struct xfs_da_args      *args,
+       struct xfs_buf          *bp,
+       int                     first,
+       int                     last)
 {
        xfs_dir2_leaf_entry_t   *firstlep;      /* pointer to first entry */
        xfs_dir2_leaf_entry_t   *lastlep;       /* pointer to last entry */
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
+       struct xfs_dir2_leaf_entry *ents;
+
+       ASSERT(leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC));
 
-       leaf = bp->data;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAF1_MAGIC ||
-              INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC);
-       firstlep = &leaf->ents[first];
-       lastlep = &leaf->ents[last];
-       xfs_da_log_buf(tp, bp, (uint)((char *)firstlep - (char *)leaf),
+       ents = args->dp->d_ops->leaf_ents_p(leaf);
+       firstlep = &ents[first];
+       lastlep = &ents[last];
+       xfs_trans_log_buf(args->trans, bp,
+               (uint)((char *)firstlep - (char *)leaf),
                (uint)((char *)lastlep - (char *)leaf + sizeof(*lastlep) - 1));
 }
 
@@ -838,37 +1124,41 @@ xfs_dir2_leaf_log_ents(
  * Log the header of the leaf1 or leafn block.
  */
 void
-xfs_dir2_leaf_log_header(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_dabuf_t             *bp)            /* leaf buffer */
+xfs_dir3_leaf_log_header(
+       struct xfs_da_args      *args,
+       struct xfs_buf          *bp)
 {
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
+
+       ASSERT(leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC));
 
-       leaf = bp->data;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAF1_MAGIC ||
-              INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC);
-       xfs_da_log_buf(tp, bp, (uint)((char *)&leaf->hdr - (char *)leaf),
-               (uint)(sizeof(leaf->hdr) - 1));
+       xfs_trans_log_buf(args->trans, bp,
+                         (uint)((char *)&leaf->hdr - (char *)leaf),
+                         args->dp->d_ops->leaf_hdr_size - 1);
 }
 
 /*
  * Log the tail of the leaf1 block.
  */
-static void
-xfs_dir2_leaf_log_tail(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_dabuf_t             *bp)            /* leaf buffer */
+STATIC void
+xfs_dir3_leaf_log_tail(
+       struct xfs_da_args      *args,
+       struct xfs_buf          *bp)
 {
-       xfs_dir2_leaf_t         *leaf;          /* leaf structure */
+       struct xfs_dir2_leaf    *leaf = bp->b_addr;
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail structure */
-       xfs_mount_t             *mp;            /* filesystem mount point */
 
-       mp = tp->t_mountp;
-       leaf = bp->data;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAF1_MAGIC);
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       xfs_da_log_buf(tp, bp, (uint)((char *)ltp - (char *)leaf),
-               (uint)(mp->m_dirblksize - 1));
+       ASSERT(leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAF1_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
+              leaf->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC));
+
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       xfs_trans_log_buf(args->trans, bp, (uint)((char *)ltp - (char *)leaf),
+               (uint)(args->geo->blksize - 1));
 }
 
 /*
@@ -880,17 +1170,19 @@ int
 xfs_dir2_leaf_lookup(
        xfs_da_args_t           *args)          /* operation arguments */
 {
-       xfs_dabuf_t             *dbp;           /* data block buffer */
+       struct xfs_buf          *dbp;           /* data block buffer */
        xfs_dir2_data_entry_t   *dep;           /* data block entry */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
        int                     index;          /* found entry index */
-       xfs_dabuf_t             *lbp;           /* leaf buffer */
+       struct xfs_buf          *lbp;           /* leaf buffer */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir2_leaf_entry *ents;
+
+       trace_xfs_dir2_leaf_lookup(args);
 
-       xfs_dir2_trace_args("leaf_lookup", args);
        /*
         * Look up name in the leaf block, returning both buffers and index.
         */
@@ -899,25 +1191,29 @@ xfs_dir2_leaf_lookup(
        }
        tp = args->trans;
        dp = args->dp;
-       xfs_dir2_leaf_check(dp, lbp);
-       leaf = lbp->data;
+       xfs_dir3_leaf_check(dp, lbp);
+       leaf = lbp->b_addr;
+       ents = dp->d_ops->leaf_ents_p(leaf);
        /*
         * Get to the leaf entry and contained data entry address.
         */
-       lep = &leaf->ents[index];
+       lep = &ents[index];
+
        /*
         * Point to the data entry.
         */
        dep = (xfs_dir2_data_entry_t *)
-             ((char *)dbp->data +
-              XFS_DIR2_DATAPTR_TO_OFF(dp->i_mount, INT_GET(lep->address, ARCH_CONVERT)));
+             ((char *)dbp->b_addr +
+              xfs_dir2_dataptr_to_off(args->geo, be32_to_cpu(lep->address)));
        /*
-        * Return the found inode number.
+        * Return the found inode number & CI name if appropriate
         */
-       args->inumber = INT_GET(dep->inumber, ARCH_CONVERT);
-       xfs_da_brelse(tp, dbp);
-       xfs_da_brelse(tp, lbp);
-       return XFS_ERROR(EEXIST);
+       args->inumber = be64_to_cpu(dep->inumber);
+       args->filetype = dp->d_ops->data_get_ftype(dep);
+       error = xfs_dir_cilookup_result(args, dep->name, dep->namelen);
+       xfs_trans_brelse(tp, dbp);
+       xfs_trans_brelse(tp, lbp);
+       return error;
 }
 
 /*
@@ -926,40 +1222,44 @@ xfs_dir2_leaf_lookup(
  * If not found dbpp will be NULL, and ENOENT comes back.
  * lbpp will always be filled in with the leaf buffer unless there's an error.
  */
-STATIC int                                     /* error */
+static int                                     /* error */
 xfs_dir2_leaf_lookup_int(
        xfs_da_args_t           *args,          /* operation arguments */
-       xfs_dabuf_t             **lbpp,         /* out: leaf buffer */
+       struct xfs_buf          **lbpp,         /* out: leaf buffer */
        int                     *indexp,        /* out: index in leaf block */
-       xfs_dabuf_t             **dbpp)         /* out: data buffer */
+       struct xfs_buf          **dbpp)         /* out: data buffer */
 {
-       xfs_dir2_db_t           curdb;          /* current data block number */
-       xfs_dabuf_t             *dbp;           /* data buffer */
+       xfs_dir2_db_t           curdb = -1;     /* current data block number */
+       struct xfs_buf          *dbp = NULL;    /* data buffer */
        xfs_dir2_data_entry_t   *dep;           /* data entry */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
        int                     index;          /* index in leaf block */
-       xfs_dabuf_t             *lbp;           /* leaf buffer */
+       struct xfs_buf          *lbp;           /* leaf buffer */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_mount_t             *mp;            /* filesystem mount point */
        xfs_dir2_db_t           newdb;          /* new data block number */
        xfs_trans_t             *tp;            /* transaction pointer */
+       xfs_dir2_db_t           cidb = -1;      /* case match data block no. */
+       enum xfs_dacmp          cmp;            /* name compare result */
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_dir3_icleaf_hdr leafhdr;
 
        dp = args->dp;
        tp = args->trans;
        mp = dp->i_mount;
-       /*
-        * Read the leaf block into the buffer.
-        */
-       if ((error =
-           xfs_da_read_buf(tp, dp, mp->m_dirleafblk, -1, &lbp,
-                   XFS_DATA_FORK))) {
+
+       error = xfs_dir3_leaf_read(tp, dp, args->geo->leafblk, -1, &lbp);
+       if (error)
                return error;
-       }
+
        *lbpp = lbp;
-       leaf = lbp->data;
-       xfs_dir2_leaf_check(dp, lbp);
+       leaf = lbp->b_addr;
+       xfs_dir3_leaf_check(dp, lbp);
+       ents = dp->d_ops->leaf_ents_p(leaf);
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+
        /*
         * Look for the first leaf entry with our hash value.
         */
@@ -968,60 +1268,87 @@ xfs_dir2_leaf_lookup_int(
         * Loop over all the entries with the right hash value
         * looking to match the name.
         */
-       for (lep = &leaf->ents[index], dbp = NULL, curdb = -1;
-            index < INT_GET(leaf->hdr.count, ARCH_CONVERT) && INT_GET(lep->hashval, ARCH_CONVERT) == args->hashval;
+       for (lep = &ents[index];
+            index < leafhdr.count && be32_to_cpu(lep->hashval) == args->hashval;
             lep++, index++) {
                /*
                 * Skip over stale leaf entries.
                 */
-               if (INT_GET(lep->address, ARCH_CONVERT) == XFS_DIR2_NULL_DATAPTR)
+               if (be32_to_cpu(lep->address) == XFS_DIR2_NULL_DATAPTR)
                        continue;
                /*
                 * Get the new data block number.
                 */
-               newdb = XFS_DIR2_DATAPTR_TO_DB(mp, INT_GET(lep->address, ARCH_CONVERT));
+               newdb = xfs_dir2_dataptr_to_db(args->geo,
+                                              be32_to_cpu(lep->address));
                /*
                 * If it's not the same as the old data block number,
                 * need to pitch the old one and read the new one.
                 */
                if (newdb != curdb) {
                        if (dbp)
-                               xfs_da_brelse(tp, dbp);
-                       if ((error =
-                           xfs_da_read_buf(tp, dp,
-                                   XFS_DIR2_DB_TO_DA(mp, newdb), -1, &dbp,
-                                   XFS_DATA_FORK))) {
-                               xfs_da_brelse(tp, lbp);
+                               xfs_trans_brelse(tp, dbp);
+                       error = xfs_dir3_data_read(tp, dp,
+                                          xfs_dir2_db_to_da(args->geo, newdb),
+                                          -1, &dbp);
+                       if (error) {
+                               xfs_trans_brelse(tp, lbp);
                                return error;
                        }
-                       xfs_dir2_data_check(dp, dbp);
                        curdb = newdb;
                }
                /*
                 * Point to the data entry.
                 */
-               dep = (xfs_dir2_data_entry_t *)
-                     ((char *)dbp->data +
-                      XFS_DIR2_DATAPTR_TO_OFF(mp, INT_GET(lep->address, ARCH_CONVERT)));
+               dep = (xfs_dir2_data_entry_t *)((char *)dbp->b_addr +
+                       xfs_dir2_dataptr_to_off(args->geo,
+                                               be32_to_cpu(lep->address)));
                /*
-                * If it matches then return it.
+                * Compare name and if it's an exact match, return the index
+                * and buffer. If it's the first case-insensitive match, store
+                * the index and buffer and continue looking for an exact match.
                 */
-               if (dep->namelen == args->namelen &&
-                   dep->name[0] == args->name[0] &&
-                   memcmp(dep->name, args->name, args->namelen) == 0) {
-                       *dbpp = dbp;
+               cmp = mp->m_dirnameops->compname(args, dep->name, dep->namelen);
+               if (cmp != XFS_CMP_DIFFERENT && cmp != args->cmpresult) {
+                       args->cmpresult = cmp;
                        *indexp = index;
-                       return 0;
+                       /* case exact match: return the current buffer. */
+                       if (cmp == XFS_CMP_EXACT) {
+                               *dbpp = dbp;
+                               return 0;
+                       }
+                       cidb = curdb;
                }
        }
+       ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
+       /*
+        * Here, we can only be doing a lookup (not a rename or remove).
+        * If a case-insensitive match was found earlier, re-read the
+        * appropriate data block if required and return it.
+        */
+       if (args->cmpresult == XFS_CMP_CASE) {
+               ASSERT(cidb != -1);
+               if (cidb != curdb) {
+                       xfs_trans_brelse(tp, dbp);
+                       error = xfs_dir3_data_read(tp, dp,
+                                          xfs_dir2_db_to_da(args->geo, cidb),
+                                          -1, &dbp);
+                       if (error) {
+                               xfs_trans_brelse(tp, lbp);
+                               return error;
+                       }
+               }
+               *dbpp = dbp;
+               return 0;
+       }
        /*
-        * No match found, return ENOENT.
+        * No match found, return -ENOENT.
         */
-       ASSERT(args->oknoent);
+       ASSERT(cidb == -1);
        if (dbp)
-               xfs_da_brelse(tp, dbp);
-       xfs_da_brelse(tp, lbp);
-       return XFS_ERROR(ENOENT);
+               xfs_trans_brelse(tp, dbp);
+       xfs_trans_brelse(tp, lbp);
+       return -ENOENT;
 }
 
 /*
@@ -1031,26 +1358,28 @@ int                                             /* error */
 xfs_dir2_leaf_removename(
        xfs_da_args_t           *args)          /* operation arguments */
 {
-       xfs_dir2_data_off_t     *bestsp;        /* leaf block best freespace */
-       xfs_dir2_data_t         *data;          /* data block structure */
+       __be16                  *bestsp;        /* leaf block best freespace */
+       xfs_dir2_data_hdr_t     *hdr;           /* data block header */
        xfs_dir2_db_t           db;             /* data block number */
-       xfs_dabuf_t             *dbp;           /* data block buffer */
+       struct xfs_buf          *dbp;           /* data block buffer */
        xfs_dir2_data_entry_t   *dep;           /* data entry structure */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
        xfs_dir2_db_t           i;              /* temporary data block # */
        int                     index;          /* index into leaf entries */
-       xfs_dabuf_t             *lbp;           /* leaf buffer */
+       struct xfs_buf          *lbp;           /* leaf buffer */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry */
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail structure */
-       xfs_mount_t             *mp;            /* filesystem mount point */
        int                     needlog;        /* need to log data header */
        int                     needscan;       /* need to rescan data frees */
        xfs_dir2_data_off_t     oldbest;        /* old value of best free */
-       xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir2_data_free *bf;          /* bestfree table */
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+
+       trace_xfs_dir2_leaf_removename(args);
 
-       xfs_dir2_trace_args("leaf_removename", args);
        /*
         * Lookup the leaf entry, get the leaf and data blocks read in.
         */
@@ -1058,59 +1387,64 @@ xfs_dir2_leaf_removename(
                return error;
        }
        dp = args->dp;
-       tp = args->trans;
-       mp = dp->i_mount;
-       leaf = lbp->data;
-       data = dbp->data;
-       xfs_dir2_data_check(dp, dbp);
+       leaf = lbp->b_addr;
+       hdr = dbp->b_addr;
+       xfs_dir3_data_check(dp, dbp);
+       bf = dp->d_ops->data_bestfree_p(hdr);
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+       ents = dp->d_ops->leaf_ents_p(leaf);
        /*
         * Point to the leaf entry, use that to point to the data entry.
         */
-       lep = &leaf->ents[index];
-       db = XFS_DIR2_DATAPTR_TO_DB(mp, INT_GET(lep->address, ARCH_CONVERT));
-       dep = (xfs_dir2_data_entry_t *)
-             ((char *)data + XFS_DIR2_DATAPTR_TO_OFF(mp, INT_GET(lep->address, ARCH_CONVERT)));
+       lep = &ents[index];
+       db = xfs_dir2_dataptr_to_db(args->geo, be32_to_cpu(lep->address));
+       dep = (xfs_dir2_data_entry_t *)((char *)hdr +
+               xfs_dir2_dataptr_to_off(args->geo, be32_to_cpu(lep->address)));
        needscan = needlog = 0;
-       oldbest = INT_GET(data->hdr.bestfree[0].length, ARCH_CONVERT);
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       bestsp = XFS_DIR2_LEAF_BESTS_P(ltp);
-       ASSERT(INT_GET(bestsp[db], ARCH_CONVERT) == oldbest);
+       oldbest = be16_to_cpu(bf[0].length);
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       bestsp = xfs_dir2_leaf_bests_p(ltp);
+       if (be16_to_cpu(bestsp[db]) != oldbest)
+               return -EFSCORRUPTED;
        /*
         * Mark the former data entry unused.
         */
-       xfs_dir2_data_make_free(tp, dbp,
-               (xfs_dir2_data_aoff_t)((char *)dep - (char *)data),
-               XFS_DIR2_DATA_ENTSIZE(dep->namelen), &needlog, &needscan);
+       xfs_dir2_data_make_free(args, dbp,
+               (xfs_dir2_data_aoff_t)((char *)dep - (char *)hdr),
+               dp->d_ops->data_entsize(dep->namelen), &needlog, &needscan);
        /*
         * We just mark the leaf entry stale by putting a null in it.
         */
-       INT_MOD(leaf->hdr.stale, ARCH_CONVERT, +1);
-       xfs_dir2_leaf_log_header(tp, lbp);
-       INT_SET(lep->address, ARCH_CONVERT, XFS_DIR2_NULL_DATAPTR);
-       xfs_dir2_leaf_log_ents(tp, lbp, index, index);
+       leafhdr.stale++;
+       dp->d_ops->leaf_hdr_to_disk(leaf, &leafhdr);
+       xfs_dir3_leaf_log_header(args, lbp);
+
+       lep->address = cpu_to_be32(XFS_DIR2_NULL_DATAPTR);
+       xfs_dir3_leaf_log_ents(args, lbp, index, index);
+
        /*
         * Scan the freespace in the data block again if necessary,
         * log the data block header if necessary.
         */
        if (needscan)
-               xfs_dir2_data_freescan(mp, data, &needlog, NULL);
+               xfs_dir2_data_freescan(dp, hdr, &needlog);
        if (needlog)
-               xfs_dir2_data_log_header(tp, dbp);
+               xfs_dir2_data_log_header(args, dbp);
        /*
         * If the longest freespace in the data block has changed,
         * put the new value in the bests table and log that.
         */
-       if (INT_GET(data->hdr.bestfree[0].length, ARCH_CONVERT) != oldbest) {
-               INT_COPY(bestsp[db], data->hdr.bestfree[0].length, ARCH_CONVERT);
-               xfs_dir2_leaf_log_bests(tp, lbp, db, db);
+       if (be16_to_cpu(bf[0].length) != oldbest) {
+               bestsp[db] = bf[0].length;
+               xfs_dir3_leaf_log_bests(args, lbp, db, db);
        }
-       xfs_dir2_data_check(dp, dbp);
+       xfs_dir3_data_check(dp, dbp);
        /*
         * If the data block is now empty then get rid of the data block.
         */
-       if (INT_GET(data->hdr.bestfree[0].length, ARCH_CONVERT) ==
-           mp->m_dirblksize - (uint)sizeof(data->hdr)) {
-               ASSERT(db != mp->m_dirdatablk);
+       if (be16_to_cpu(bf[0].length) ==
+                       args->geo->blksize - dp->d_ops->data_entry_offset) {
+               ASSERT(db != args->geo->datablk);
                if ((error = xfs_dir2_shrink_inode(args, db, dbp))) {
                        /*
                         * Nope, can't get rid of it because it caused
@@ -1118,12 +1452,9 @@ xfs_dir2_leaf_removename(
                         * Just go on, returning success, leaving the
                         * empty block in place.
                         */
-                       if (error == ENOSPC && args->total == 0) {
-                               xfs_da_buf_done(dbp);
+                       if (error == -ENOSPC && args->total == 0)
                                error = 0;
-                       }
-                       xfs_dir2_leaf_check(dp, lbp);
-                       xfs_da_buf_done(lbp);
+                       xfs_dir3_leaf_check(dp, lbp);
                        return error;
                }
                dbp = NULL;
@@ -1131,12 +1462,12 @@ xfs_dir2_leaf_removename(
                 * If this is the last data block then compact the
                 * bests table by getting rid of entries.
                 */
-               if (db == INT_GET(ltp->bestcount, ARCH_CONVERT) - 1) {
+               if (db == be32_to_cpu(ltp->bestcount) - 1) {
                        /*
                         * Look for the last active entry (i).
                         */
                        for (i = db - 1; i > 0; i--) {
-                               if (INT_GET(bestsp[i], ARCH_CONVERT) != NULLDATAOFF)
+                               if (bestsp[i] != cpu_to_be16(NULLDATAOFF))
                                        break;
                        }
                        /*
@@ -1144,21 +1475,21 @@ xfs_dir2_leaf_removename(
                         * end are removed.
                         */
                        memmove(&bestsp[db - i], bestsp,
-                               (INT_GET(ltp->bestcount, ARCH_CONVERT) - (db - i)) * sizeof(*bestsp));
-                       INT_MOD(ltp->bestcount, ARCH_CONVERT, -(db - i));
-                       xfs_dir2_leaf_log_tail(tp, lbp);
-                       xfs_dir2_leaf_log_bests(tp, lbp, 0, INT_GET(ltp->bestcount, ARCH_CONVERT) - 1);
+                               (be32_to_cpu(ltp->bestcount) - (db - i)) * sizeof(*bestsp));
+                       be32_add_cpu(&ltp->bestcount, -(db - i));
+                       xfs_dir3_leaf_log_tail(args, lbp);
+                       xfs_dir3_leaf_log_bests(args, lbp, 0,
+                                               be32_to_cpu(ltp->bestcount) - 1);
                } else
-                       INT_SET(bestsp[db], ARCH_CONVERT, NULLDATAOFF);
+                       bestsp[db] = cpu_to_be16(NULLDATAOFF);
        }
        /*
         * If the data block was not the first one, drop it.
         */
-       else if (db != mp->m_dirdatablk && dbp != NULL) {
-               xfs_da_buf_done(dbp);
+       else if (db != args->geo->datablk)
                dbp = NULL;
-       }
-       xfs_dir2_leaf_check(dp, lbp);
+
+       xfs_dir3_leaf_check(dp, lbp);
        /*
         * See if we can convert to block form.
         */
@@ -1172,17 +1503,19 @@ int                                             /* error */
 xfs_dir2_leaf_replace(
        xfs_da_args_t           *args)          /* operation arguments */
 {
-       xfs_dabuf_t             *dbp;           /* data block buffer */
+       struct xfs_buf          *dbp;           /* data block buffer */
        xfs_dir2_data_entry_t   *dep;           /* data block entry */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
        int                     index;          /* index of leaf entry */
-       xfs_dabuf_t             *lbp;           /* leaf buffer */
+       struct xfs_buf          *lbp;           /* leaf buffer */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir2_leaf_entry *ents;
+
+       trace_xfs_dir2_leaf_replace(args);
 
-       xfs_dir2_trace_args("leaf_replace", args);
        /*
         * Look up the entry.
         */
@@ -1190,27 +1523,28 @@ xfs_dir2_leaf_replace(
                return error;
        }
        dp = args->dp;
-       leaf = lbp->data;
+       leaf = lbp->b_addr;
+       ents = dp->d_ops->leaf_ents_p(leaf);
        /*
         * Point to the leaf entry, get data address from it.
         */
-       lep = &leaf->ents[index];
+       lep = &ents[index];
        /*
         * Point to the data entry.
         */
        dep = (xfs_dir2_data_entry_t *)
-             ((char *)dbp->data +
-              XFS_DIR2_DATAPTR_TO_OFF(dp->i_mount, INT_GET(lep->address, ARCH_CONVERT)));
-       ASSERT(args->inumber != INT_GET(dep->inumber, ARCH_CONVERT));
+             ((char *)dbp->b_addr +
+              xfs_dir2_dataptr_to_off(args->geo, be32_to_cpu(lep->address)));
+       ASSERT(args->inumber != be64_to_cpu(dep->inumber));
        /*
         * Put the new inode number in, log it.
         */
-       INT_SET(dep->inumber, ARCH_CONVERT, args->inumber);
+       dep->inumber = cpu_to_be64(args->inumber);
+       dp->d_ops->data_put_ftype(dep, args->filetype);
        tp = args->trans;
-       xfs_dir2_data_log_entry(tp, dbp, dep);
-       xfs_da_buf_done(dbp);
-       xfs_dir2_leaf_check(dp, lbp);
-       xfs_da_brelse(tp, lbp);
+       xfs_dir2_data_log_entry(args, dbp, dep);
+       xfs_dir3_leaf_check(dp, lbp);
+       xfs_trans_brelse(tp, lbp);
        return 0;
 }
 
@@ -1222,7 +1556,7 @@ xfs_dir2_leaf_replace(
 int                                            /* index value */
 xfs_dir2_leaf_search_hash(
        xfs_da_args_t           *args,          /* operation arguments */
-       xfs_dabuf_t             *lbp)           /* leaf buffer */
+       struct xfs_buf          *lbp)           /* leaf buffer */
 {
        xfs_dahash_t            hash=0;         /* hash from this entry */
        xfs_dahash_t            hashwant;       /* hash value looking for */
@@ -1231,21 +1565,22 @@ xfs_dir2_leaf_search_hash(
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_entry_t   *lep;           /* leaf entry */
        int                     mid=0;          /* current leaf index */
+       struct xfs_dir2_leaf_entry *ents;
+       struct xfs_dir3_icleaf_hdr leafhdr;
+
+       leaf = lbp->b_addr;
+       ents = args->dp->d_ops->leaf_ents_p(leaf);
+       args->dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
 
-       leaf = lbp->data;
-#ifndef __KERNEL__
-       if (!leaf->hdr.count)
-               return 0;
-#endif
        /*
         * Note, the table cannot be empty, so we have to go through the loop.
         * Binary search the leaf entries looking for our hash value.
         */
-       for (lep = leaf->ents, low = 0, high = INT_GET(leaf->hdr.count, ARCH_CONVERT) - 1,
+       for (lep = ents, low = 0, high = leafhdr.count - 1,
                hashwant = args->hashval;
             low <= high; ) {
                mid = (low + high) >> 1;
-               if ((hash = INT_GET(lep[mid].hashval, ARCH_CONVERT)) == hashwant)
+               if ((hash = be32_to_cpu(lep[mid].hashval)) == hashwant)
                        break;
                if (hash < hashwant)
                        low = mid + 1;
@@ -1256,7 +1591,7 @@ xfs_dir2_leaf_search_hash(
         * Found one, back up through all the equal hash values.
         */
        if (hash == hashwant) {
-               while (mid > 0 && INT_GET(lep[mid - 1].hashval, ARCH_CONVERT) == hashwant) {
+               while (mid > 0 && be32_to_cpu(lep[mid - 1].hashval) == hashwant) {
                        mid--;
                }
        }
@@ -1275,64 +1610,82 @@ xfs_dir2_leaf_search_hash(
 int                                            /* error */
 xfs_dir2_leaf_trim_data(
        xfs_da_args_t           *args,          /* operation arguments */
-       xfs_dabuf_t             *lbp,           /* leaf buffer */
+       struct xfs_buf          *lbp,           /* leaf buffer */
        xfs_dir2_db_t           db)             /* data block number */
 {
-       xfs_dir2_data_off_t     *bestsp;        /* leaf bests table */
-#ifdef DEBUG
-       xfs_dir2_data_t         *data;          /* data block structure */
-#endif
-       xfs_dabuf_t             *dbp;           /* data block buffer */
+       __be16                  *bestsp;        /* leaf bests table */
+       struct xfs_buf          *dbp;           /* data block buffer */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return value */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_dir2_leaf_tail_t    *ltp;           /* leaf tail structure */
-       xfs_mount_t             *mp;            /* filesystem mount point */
        xfs_trans_t             *tp;            /* transaction pointer */
 
        dp = args->dp;
-       mp = dp->i_mount;
        tp = args->trans;
        /*
         * Read the offending data block.  We need its buffer.
         */
-       if ((error = xfs_da_read_buf(tp, dp, XFS_DIR2_DB_TO_DA(mp, db), -1, &dbp,
-                       XFS_DATA_FORK))) {
+       error = xfs_dir3_data_read(tp, dp, xfs_dir2_db_to_da(args->geo, db),
+                                  -1, &dbp);
+       if (error)
                return error;
-       }
+
+       leaf = lbp->b_addr;
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+
 #ifdef DEBUG
-       data = dbp->data;
-       ASSERT(INT_GET(data->hdr.magic, ARCH_CONVERT) == XFS_DIR2_DATA_MAGIC);
+{
+       struct xfs_dir2_data_hdr *hdr = dbp->b_addr;
+       struct xfs_dir2_data_free *bf = dp->d_ops->data_bestfree_p(hdr);
+
+       ASSERT(hdr->magic == cpu_to_be32(XFS_DIR2_DATA_MAGIC) ||
+              hdr->magic == cpu_to_be32(XFS_DIR3_DATA_MAGIC));
+       ASSERT(be16_to_cpu(bf[0].length) ==
+              args->geo->blksize - dp->d_ops->data_entry_offset);
+       ASSERT(db == be32_to_cpu(ltp->bestcount) - 1);
+}
 #endif
-       /* this seems to be an error
-        * data is only valid if DEBUG is defined?
-        * RMC 09/08/1999
-        */
 
-       leaf = lbp->data;
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       ASSERT(INT_GET(data->hdr.bestfree[0].length, ARCH_CONVERT) ==
-              mp->m_dirblksize - (uint)sizeof(data->hdr));
-       ASSERT(db == INT_GET(ltp->bestcount, ARCH_CONVERT) - 1);
        /*
         * Get rid of the data block.
         */
        if ((error = xfs_dir2_shrink_inode(args, db, dbp))) {
-               ASSERT(error != ENOSPC);
-               xfs_da_brelse(tp, dbp);
+               ASSERT(error != -ENOSPC);
+               xfs_trans_brelse(tp, dbp);
                return error;
        }
        /*
         * Eliminate the last bests entry from the table.
         */
-       bestsp = XFS_DIR2_LEAF_BESTS_P(ltp);
-       INT_MOD(ltp->bestcount, ARCH_CONVERT, -1);
-       memmove(&bestsp[1], &bestsp[0], INT_GET(ltp->bestcount, ARCH_CONVERT) * sizeof(*bestsp));
-       xfs_dir2_leaf_log_tail(tp, lbp);
-       xfs_dir2_leaf_log_bests(tp, lbp, 0, INT_GET(ltp->bestcount, ARCH_CONVERT) - 1);
+       bestsp = xfs_dir2_leaf_bests_p(ltp);
+       be32_add_cpu(&ltp->bestcount, -1);
+       memmove(&bestsp[1], &bestsp[0], be32_to_cpu(ltp->bestcount) * sizeof(*bestsp));
+       xfs_dir3_leaf_log_tail(args, lbp);
+       xfs_dir3_leaf_log_bests(args, lbp, 0, be32_to_cpu(ltp->bestcount) - 1);
        return 0;
 }
 
+static inline size_t
+xfs_dir3_leaf_size(
+       struct xfs_dir3_icleaf_hdr      *hdr,
+       int                             counts)
+{
+       int     entries;
+       int     hdrsize;
+
+       entries = hdr->count - hdr->stale;
+       if (hdr->magic == XFS_DIR2_LEAF1_MAGIC ||
+           hdr->magic == XFS_DIR2_LEAFN_MAGIC)
+               hdrsize = sizeof(struct xfs_dir2_leaf_hdr);
+       else
+               hdrsize = sizeof(struct xfs_dir3_leaf_hdr);
+
+       return hdrsize + entries * sizeof(xfs_dir2_leaf_entry_t)
+                      + counts * sizeof(xfs_dir2_data_off_t)
+                      + sizeof(xfs_dir2_leaf_tail_t);
+}
+
 /*
  * Convert node form directory to leaf form directory.
  * The root of the node form dir needs to already be a LEAFN block.
@@ -1345,15 +1698,17 @@ xfs_dir2_node_to_leaf(
        xfs_da_args_t           *args;          /* operation arguments */
        xfs_inode_t             *dp;            /* incore directory inode */
        int                     error;          /* error return code */
-       xfs_dabuf_t             *fbp;           /* buffer for freespace block */
+       struct xfs_buf          *fbp;           /* buffer for freespace block */
        xfs_fileoff_t           fo;             /* freespace file offset */
        xfs_dir2_free_t         *free;          /* freespace structure */
-       xfs_dabuf_t             *lbp;           /* buffer for leaf block */
+       struct xfs_buf          *lbp;           /* buffer for leaf block */
        xfs_dir2_leaf_tail_t    *ltp;           /* tail of leaf structure */
        xfs_dir2_leaf_t         *leaf;          /* leaf structure */
        xfs_mount_t             *mp;            /* filesystem mount point */
        int                     rval;           /* successful free trim? */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir3_icleaf_hdr leafhdr;
+       struct xfs_dir3_icfree_hdr freehdr;
 
        /*
         * There's more than a leaf level in the btree, so there must
@@ -1362,29 +1717,31 @@ xfs_dir2_node_to_leaf(
        if (state->path.active > 1)
                return 0;
        args = state->args;
-       xfs_dir2_trace_args("node_to_leaf", args);
+
+       trace_xfs_dir2_node_to_leaf(args);
+
        mp = state->mp;
        dp = args->dp;
        tp = args->trans;
        /*
         * Get the last offset in the file.
         */
-       if ((error = xfs_bmap_last_offset(tp, dp, &fo, XFS_DATA_FORK))) {
+       if ((error = xfs_bmap_last_offset(dp, &fo, XFS_DATA_FORK))) {
                return error;
        }
-       fo -= mp->m_dirblkfsbs;
+       fo -= args->geo->fsbcount;
        /*
         * If there are freespace blocks other than the first one,
         * take this opportunity to remove trailing empty freespace blocks
         * that may have been left behind during no-space-reservation
         * operations.
         */
-       while (fo > mp->m_dirfreeblk) {
+       while (fo > args->geo->freeblk) {
                if ((error = xfs_dir2_node_trim_free(args, fo, &rval))) {
                        return error;
                }
                if (rval)
-                       fo -= mp->m_dirblkfsbs;
+                       fo -= args->geo->fsbcount;
                else
                        return 0;
        }
@@ -1397,66 +1754,78 @@ xfs_dir2_node_to_leaf(
        /*
         * If it's not the single leaf block, give up.
         */
-       if (XFS_FSB_TO_B(mp, fo) > XFS_DIR2_LEAF_OFFSET + mp->m_dirblksize)
+       if (XFS_FSB_TO_B(mp, fo) > XFS_DIR2_LEAF_OFFSET + args->geo->blksize)
                return 0;
        lbp = state->path.blk[0].bp;
-       leaf = lbp->data;
-       ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC);
+       leaf = lbp->b_addr;
+       dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
+
+       ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
+              leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
+
        /*
         * Read the freespace block.
         */
-       if ((error = xfs_da_read_buf(tp, dp, mp->m_dirfreeblk, -1, &fbp,
-                       XFS_DATA_FORK))) {
+       error = xfs_dir2_free_read(tp, dp,  args->geo->freeblk, &fbp);
+       if (error)
                return error;
-       }
-       free = fbp->data;
-       ASSERT(INT_GET(free->hdr.magic, ARCH_CONVERT) == XFS_DIR2_FREE_MAGIC);
-       ASSERT(!free->hdr.firstdb);
+       free = fbp->b_addr;
+       dp->d_ops->free_hdr_from_disk(&freehdr, free);
+
+       ASSERT(!freehdr.firstdb);
+
        /*
         * Now see if the leafn and free data will fit in a leaf1.
         * If not, release the buffer and give up.
         */
-       if ((uint)sizeof(leaf->hdr) +
-           (INT_GET(leaf->hdr.count, ARCH_CONVERT) - INT_GET(leaf->hdr.stale, ARCH_CONVERT)) * (uint)sizeof(leaf->ents[0]) +
-           INT_GET(free->hdr.nvalid, ARCH_CONVERT) * (uint)sizeof(leaf->bests[0]) +
-           (uint)sizeof(leaf->tail) >
-           mp->m_dirblksize) {
-               xfs_da_brelse(tp, fbp);
+       if (xfs_dir3_leaf_size(&leafhdr, freehdr.nvalid) > args->geo->blksize) {
+               xfs_trans_brelse(tp, fbp);
                return 0;
        }
+
        /*
         * If the leaf has any stale entries in it, compress them out.
-        * The compact routine will log the header.
         */
-       if (INT_GET(leaf->hdr.stale, ARCH_CONVERT))
-               xfs_dir2_leaf_compact(args, lbp);
-       else
-               xfs_dir2_leaf_log_header(tp, lbp);
-       INT_SET(leaf->hdr.info.magic, ARCH_CONVERT, XFS_DIR2_LEAF1_MAGIC);
+       if (leafhdr.stale)
+               xfs_dir3_leaf_compact(args, &leafhdr, lbp);
+
+       lbp->b_ops = &xfs_dir3_leaf1_buf_ops;
+       xfs_trans_buf_set_type(tp, lbp, XFS_BLFT_DIR_LEAF1_BUF);
+       leafhdr.magic = (leafhdr.magic == XFS_DIR2_LEAFN_MAGIC)
+                                       ? XFS_DIR2_LEAF1_MAGIC
+                                       : XFS_DIR3_LEAF1_MAGIC;
+
        /*
         * Set up the leaf tail from the freespace block.
         */
-       ltp = XFS_DIR2_LEAF_TAIL_P(mp, leaf);
-       INT_COPY(ltp->bestcount, free->hdr.nvalid, ARCH_CONVERT);
+       ltp = xfs_dir2_leaf_tail_p(args->geo, leaf);
+       ltp->bestcount = cpu_to_be32(freehdr.nvalid);
+
        /*
         * Set up the leaf bests table.
         */
-       memcpy(XFS_DIR2_LEAF_BESTS_P(ltp), free->bests,
-               INT_GET(ltp->bestcount, ARCH_CONVERT) * sizeof(leaf->bests[0]));
-       xfs_dir2_leaf_log_bests(tp, lbp, 0, INT_GET(ltp->bestcount, ARCH_CONVERT) - 1);
-       xfs_dir2_leaf_log_tail(tp, lbp);
-       xfs_dir2_leaf_check(dp, lbp);
+       memcpy(xfs_dir2_leaf_bests_p(ltp), dp->d_ops->free_bests_p(free),
+               freehdr.nvalid * sizeof(xfs_dir2_data_off_t));
+
+       dp->d_ops->leaf_hdr_to_disk(leaf, &leafhdr);
+       xfs_dir3_leaf_log_header(args, lbp);
+       xfs_dir3_leaf_log_bests(args, lbp, 0, be32_to_cpu(ltp->bestcount) - 1);
+       xfs_dir3_leaf_log_tail(args, lbp);
+       xfs_dir3_leaf_check(dp, lbp);
+
        /*
         * Get rid of the freespace block.
         */
-       error = xfs_dir2_shrink_inode(args, XFS_DIR2_FREE_FIRSTDB(mp), fbp);
+       error = xfs_dir2_shrink_inode(args,
+                       xfs_dir2_byte_to_db(args->geo, XFS_DIR2_FREE_OFFSET),
+                       fbp);
        if (error) {
                /*
                 * This can't fail here because it can only happen when
                 * punching out the middle of an extent, and this is an
                 * isolated block.
                 */
-               ASSERT(error != ENOSPC);
+               ASSERT(error != -ENOSPC);
                return error;
        }
        fbp = NULL;