]> git.openfabrics.org - ~emulex/infiniband.git/commitdiff
xfs: add CRC checking to dir2 free blocks
authorDave Chinner <dchinner@redhat.com>
Wed, 3 Apr 2013 05:11:21 +0000 (16:11 +1100)
committerBen Myers <bpm@sgi.com>
Sat, 27 Apr 2013 16:58:16 +0000 (11:58 -0500)
This addition follows the same pattern as the dir2 block CRCs, but
with a few differences. The main difference is that the free block
header is different between the v2 and v3 formats, so an "in-core"
free block header has been added and _todisk/_from_disk functions
used to abstract the differences in structure format from the code.
This is similar to the on-disk superblock versus the in-core
superblock setup. The in-core strucutre is populated when the buffer
is read from disk, all the in memory checks and modifications are
done on the in-core version of the structure which is written back
to the buffer before the buffer is logged.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
fs/xfs/xfs_dir2_format.h
fs/xfs/xfs_dir2_leaf.c
fs/xfs/xfs_dir2_node.c

index 845979a4ee969737a1018787ded98230bce03384..ec5044a92b7045df9dcb32ab7d75e55961f14011 100644 (file)
@@ -66,6 +66,7 @@
 
 #define        XFS_DIR3_BLOCK_MAGIC    0x58444233      /* XDB3: single block dirs */
 #define        XFS_DIR3_DATA_MAGIC     0x58444433      /* XDD3: multiblock dirs */
+#define        XFS_DIR3_FREE_MAGIC     0x58444633      /* XDF3: free index blocks */
 
 /*
  * Byte offset in data block and shortform entry.
@@ -663,19 +664,65 @@ typedef struct xfs_dir2_free {
                                                /* unused entries are -1 */
 } xfs_dir2_free_t;
 
-static inline int xfs_dir2_free_max_bests(struct xfs_mount *mp)
+struct xfs_dir3_free_hdr {
+       struct xfs_dir3_blk_hdr hdr;
+       __be32                  firstdb;        /* db of first entry */
+       __be32                  nvalid;         /* count of valid entries */
+       __be32                  nused;          /* count of used entries */
+};
+
+struct xfs_dir3_free {
+       struct xfs_dir3_free_hdr hdr;
+       __be16                  bests[];        /* best free counts */
+                                               /* unused entries are -1 */
+};
+
+#define XFS_DIR3_FREE_CRC_OFF  offsetof(struct xfs_dir3_free, hdr.hdr.crc)
+
+/*
+ * In core version of the free block header, abstracted away from on-disk format
+ * differences. Use this in the code, and convert to/from the disk version using
+ * xfs_dir3_free_hdr_from_disk/xfs_dir3_free_hdr_to_disk.
+ */
+struct xfs_dir3_icfree_hdr {
+       __uint32_t      magic;
+       __uint32_t      firstdb;
+       __uint32_t      nvalid;
+       __uint32_t      nused;
+
+};
+
+void xfs_dir3_free_hdr_from_disk(struct xfs_dir3_icfree_hdr *to,
+                                struct xfs_dir2_free *from);
+
+static inline int
+xfs_dir3_free_hdr_size(struct xfs_mount *mp)
 {
-       return (mp->m_dirblksize - sizeof(struct xfs_dir2_free_hdr)) /
+       if (xfs_sb_version_hascrc(&mp->m_sb))
+               return sizeof(struct xfs_dir3_free_hdr);
+       return sizeof(struct xfs_dir2_free_hdr);
+}
+
+static inline int
+xfs_dir3_free_max_bests(struct xfs_mount *mp)
+{
+       return (mp->m_dirblksize - xfs_dir3_free_hdr_size(mp)) /
                sizeof(xfs_dir2_data_off_t);
 }
 
+static inline __be16 *
+xfs_dir3_free_bests_p(struct xfs_mount *mp, struct xfs_dir2_free *free)
+{
+       return (__be16 *)((char *)free + xfs_dir3_free_hdr_size(mp));
+}
+
 /*
  * Convert data space db to the corresponding free db.
  */
 static inline xfs_dir2_db_t
 xfs_dir2_db_to_fdb(struct xfs_mount *mp, xfs_dir2_db_t db)
 {
-       return XFS_DIR2_FREE_FIRSTDB(mp) + db / xfs_dir2_free_max_bests(mp);
+       return XFS_DIR2_FREE_FIRSTDB(mp) + db / xfs_dir3_free_max_bests(mp);
 }
 
 /*
@@ -684,7 +731,7 @@ xfs_dir2_db_to_fdb(struct xfs_mount *mp, xfs_dir2_db_t db)
 static inline int
 xfs_dir2_db_to_fdindex(struct xfs_mount *mp, xfs_dir2_db_t db)
 {
-       return db % xfs_dir2_free_max_bests(mp);
+       return db % xfs_dir3_free_max_bests(mp);
 }
 
 /*
index ef163f06fd16cb99db8046cb5332073cfc35c833..979735b0c76deef832ba7f5707567e5e6a88b0f5 100644 (file)
@@ -1881,6 +1881,7 @@ xfs_dir2_node_to_leaf(
        xfs_mount_t             *mp;            /* filesystem mount point */
        int                     rval;           /* successful free trim? */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir3_icfree_hdr freehdr;
 
        /*
         * There's more than a leaf level in the btree, so there must
@@ -1938,15 +1939,15 @@ xfs_dir2_node_to_leaf(
        if (error)
                return error;
        free = fbp->b_addr;
-       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
-       ASSERT(!free->hdr.firstdb);
+       xfs_dir3_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 (xfs_dir2_leaf_size(&leaf->hdr, be32_to_cpu(free->hdr.nvalid)) >
-                       mp->m_dirblksize) {
+       if (xfs_dir2_leaf_size(&leaf->hdr, freehdr.nvalid) > mp->m_dirblksize) {
                xfs_trans_brelse(tp, fbp);
                return 0;
        }
@@ -1967,12 +1968,12 @@ xfs_dir2_node_to_leaf(
         * Set up the leaf tail from the freespace block.
         */
        ltp = xfs_dir2_leaf_tail_p(mp, leaf);
-       ltp->bestcount = free->hdr.nvalid;
+       ltp->bestcount = cpu_to_be32(freehdr.nvalid);
        /*
         * Set up the leaf bests table.
         */
-       memcpy(xfs_dir2_leaf_bests_p(ltp), free->bests,
-               be32_to_cpu(ltp->bestcount) * sizeof(xfs_dir2_data_off_t));
+       memcpy(xfs_dir2_leaf_bests_p(ltp), xfs_dir3_free_bests_p(mp, free),
+               freehdr.nvalid * sizeof(xfs_dir2_data_off_t));
        xfs_dir2_leaf_log_bests(tp, lbp, 0, be32_to_cpu(ltp->bestcount) - 1);
        xfs_dir2_leaf_log_tail(tp, lbp);
        xfs_dir2_leaf_check(dp, lbp);
index 985d70bc391d91b512f2005d461c46d49d5f5b45..0fe39b9fde7eeb7dfbb4f7f7226b4d853eb31d84 100644 (file)
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
+ * Copyright (c) 2013 Red Hat, Inc.
  * All Rights Reserved.
  *
  * This program is free software; you can redistribute it and/or
@@ -32,6 +33,8 @@
 #include "xfs_dir2_priv.h"
 #include "xfs_error.h"
 #include "xfs_trace.h"
+#include "xfs_buf_item.h"
+#include "xfs_cksum.h"
 
 /*
  * Function declarations.
@@ -55,44 +58,78 @@ static int xfs_dir2_leafn_remove(xfs_da_args_t *args, struct xfs_buf *bp,
 static int xfs_dir2_node_addname_int(xfs_da_args_t *args,
                                     xfs_da_state_blk_t *fblk);
 
-static void
-xfs_dir2_free_verify(
+static bool
+xfs_dir3_free_verify(
        struct xfs_buf          *bp)
 {
        struct xfs_mount        *mp = bp->b_target->bt_mount;
        struct xfs_dir2_free_hdr *hdr = bp->b_addr;
-       int                     block_ok = 0;
 
-       block_ok = hdr->magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC);
-       if (!block_ok) {
-               XFS_CORRUPTION_ERROR("xfs_dir2_free_verify magic",
-                                    XFS_ERRLEVEL_LOW, mp, hdr);
-               xfs_buf_ioerror(bp, EFSCORRUPTED);
+       if (xfs_sb_version_hascrc(&mp->m_sb)) {
+               struct xfs_dir3_blk_hdr *hdr3 = bp->b_addr;
+
+               if (hdr3->magic != cpu_to_be32(XFS_DIR3_FREE_MAGIC))
+                       return false;
+               if (!uuid_equal(&hdr3->uuid, &mp->m_sb.sb_uuid))
+                       return false;
+               if (be64_to_cpu(hdr3->blkno) != bp->b_bn)
+                       return false;
+       } else {
+               if (hdr->magic != cpu_to_be32(XFS_DIR2_FREE_MAGIC))
+                       return false;
        }
+
+       /* XXX: should bounds check the xfs_dir3_icfree_hdr here */
+
+       return true;
 }
 
 static void
-xfs_dir2_free_read_verify(
+xfs_dir3_free_read_verify(
        struct xfs_buf  *bp)
 {
-       xfs_dir2_free_verify(bp);
+       struct xfs_mount        *mp = bp->b_target->bt_mount;
+
+       if ((xfs_sb_version_hascrc(&mp->m_sb) &&
+            !xfs_verify_cksum(bp->b_addr, BBTOB(bp->b_length),
+                                         XFS_DIR3_FREE_CRC_OFF)) ||
+           !xfs_dir3_free_verify(bp)) {
+               XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
+               xfs_buf_ioerror(bp, EFSCORRUPTED);
+       }
 }
 
 static void
-xfs_dir2_free_write_verify(
+xfs_dir3_free_write_verify(
        struct xfs_buf  *bp)
 {
-       xfs_dir2_free_verify(bp);
+       struct xfs_mount        *mp = bp->b_target->bt_mount;
+       struct xfs_buf_log_item *bip = bp->b_fspriv;
+       struct xfs_dir3_blk_hdr *hdr3 = bp->b_addr;
+
+       if (!xfs_dir3_free_verify(bp)) {
+               XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
+               xfs_buf_ioerror(bp, EFSCORRUPTED);
+               return;
+       }
+
+       if (!xfs_sb_version_hascrc(&mp->m_sb))
+               return;
+
+       if (bip)
+               hdr3->lsn = cpu_to_be64(bip->bli_item.li_lsn);
+
+       xfs_update_cksum(bp->b_addr, BBTOB(bp->b_length), XFS_DIR3_FREE_CRC_OFF);
 }
 
-static const struct xfs_buf_ops xfs_dir2_free_buf_ops = {
-       .verify_read = xfs_dir2_free_read_verify,
-       .verify_write = xfs_dir2_free_write_verify,
+static const struct xfs_buf_ops xfs_dir3_free_buf_ops = {
+       .verify_read = xfs_dir3_free_read_verify,
+       .verify_write = xfs_dir3_free_write_verify,
 };
 
 
 static int
-__xfs_dir2_free_read(
+__xfs_dir3_free_read(
        struct xfs_trans        *tp,
        struct xfs_inode        *dp,
        xfs_dablk_t             fbno,
@@ -100,7 +137,7 @@ __xfs_dir2_free_read(
        struct xfs_buf          **bpp)
 {
        return xfs_da_read_buf(tp, dp, fbno, mappedbno, bpp,
-                               XFS_DATA_FORK, &xfs_dir2_free_buf_ops);
+                               XFS_DATA_FORK, &xfs_dir3_free_buf_ops);
 }
 
 int
@@ -110,7 +147,7 @@ xfs_dir2_free_read(
        xfs_dablk_t             fbno,
        struct xfs_buf          **bpp)
 {
-       return __xfs_dir2_free_read(tp, dp, fbno, -1, bpp);
+       return __xfs_dir3_free_read(tp, dp, fbno, -1, bpp);
 }
 
 static int
@@ -120,7 +157,94 @@ xfs_dir2_free_try_read(
        xfs_dablk_t             fbno,
        struct xfs_buf          **bpp)
 {
-       return __xfs_dir2_free_read(tp, dp, fbno, -2, bpp);
+       return __xfs_dir3_free_read(tp, dp, fbno, -2, bpp);
+}
+
+
+void
+xfs_dir3_free_hdr_from_disk(
+       struct xfs_dir3_icfree_hdr      *to,
+       struct xfs_dir2_free            *from)
+{
+       if (from->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC)) {
+               to->magic = be32_to_cpu(from->hdr.magic);
+               to->firstdb = be32_to_cpu(from->hdr.firstdb);
+               to->nvalid = be32_to_cpu(from->hdr.nvalid);
+               to->nused = be32_to_cpu(from->hdr.nused);
+       } else {
+               struct xfs_dir3_free_hdr *hdr3 = (struct xfs_dir3_free_hdr *)from;
+
+               to->magic = be32_to_cpu(hdr3->hdr.magic);
+               to->firstdb = be32_to_cpu(hdr3->firstdb);
+               to->nvalid = be32_to_cpu(hdr3->nvalid);
+               to->nused = be32_to_cpu(hdr3->nused);
+       }
+
+       ASSERT(to->magic == XFS_DIR2_FREE_MAGIC ||
+              to->magic == XFS_DIR3_FREE_MAGIC);
+}
+
+static void
+xfs_dir3_free_hdr_to_disk(
+       struct xfs_dir2_free            *to,
+       struct xfs_dir3_icfree_hdr      *from)
+{
+       ASSERT(from->magic == XFS_DIR2_FREE_MAGIC ||
+              from->magic == XFS_DIR3_FREE_MAGIC);
+
+       if (from->magic == XFS_DIR2_FREE_MAGIC) {
+               to->hdr.magic = cpu_to_be32(from->magic);
+               to->hdr.firstdb = cpu_to_be32(from->firstdb);
+               to->hdr.nvalid = cpu_to_be32(from->nvalid);
+               to->hdr.nused = cpu_to_be32(from->nused);
+       } else {
+               struct xfs_dir3_free_hdr *hdr3 = (struct xfs_dir3_free_hdr *)to;
+
+               hdr3->hdr.magic = cpu_to_be32(from->magic);
+               hdr3->firstdb = cpu_to_be32(from->firstdb);
+               hdr3->nvalid = cpu_to_be32(from->nvalid);
+               hdr3->nused = cpu_to_be32(from->nused);
+       }
+}
+
+static int
+xfs_dir3_free_get_buf(
+       struct xfs_trans        *tp,
+       struct xfs_inode        *dp,
+       xfs_dir2_db_t           fbno,
+       struct xfs_buf          **bpp)
+{
+       struct xfs_mount        *mp = dp->i_mount;
+       struct xfs_buf          *bp;
+       int                     error;
+       struct xfs_dir3_icfree_hdr hdr;
+
+       error = xfs_da_get_buf(tp, dp, xfs_dir2_db_to_da(mp, fbno),
+                                  -1, &bp, XFS_DATA_FORK);
+       if (error)
+               return error;
+
+       bp->b_ops = &xfs_dir3_free_buf_ops;
+
+       /*
+        * Initialize the new block to be empty, and remember
+        * its first slot as our empty slot.
+        */
+       hdr.magic = XFS_DIR2_FREE_MAGIC;
+       hdr.firstdb = 0;
+       hdr.nused = 0;
+       hdr.nvalid = 0;
+       if (xfs_sb_version_hascrc(&mp->m_sb)) {
+               struct xfs_dir3_free_hdr *hdr3 = bp->b_addr;
+
+               hdr.magic = XFS_DIR3_FREE_MAGIC;
+               hdr3->hdr.blkno = cpu_to_be64(bp->b_bn);
+               hdr3->hdr.owner = cpu_to_be64(dp->i_ino);
+               uuid_copy(&hdr3->hdr.uuid, &mp->m_sb.sb_uuid);
+       }
+       xfs_dir3_free_hdr_to_disk(bp->b_addr, &hdr);
+       *bpp = bp;
+       return 0;
 }
 
 /*
@@ -134,13 +258,16 @@ xfs_dir2_free_log_bests(
        int                     last)           /* last entry to log */
 {
        xfs_dir2_free_t         *free;          /* freespace structure */
+       __be16                  *bests;
 
        free = bp->b_addr;
-       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
+       bests = xfs_dir3_free_bests_p(tp->t_mountp, free);
+       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC) ||
+              free->hdr.magic == cpu_to_be32(XFS_DIR3_FREE_MAGIC));
        xfs_trans_log_buf(tp, bp,
-               (uint)((char *)&free->bests[first] - (char *)free),
-               (uint)((char *)&free->bests[last] - (char *)free +
-                      sizeof(free->bests[0]) - 1));
+               (uint)((char *)&bests[first] - (char *)free),
+               (uint)((char *)&bests[last] - (char *)free +
+                      sizeof(bests[0]) - 1));
 }
 
 /*
@@ -154,9 +281,9 @@ xfs_dir2_free_log_header(
        xfs_dir2_free_t         *free;          /* freespace structure */
 
        free = bp->b_addr;
-       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
-       xfs_trans_log_buf(tp, bp, (uint)((char *)&free->hdr - (char *)free),
-               (uint)(sizeof(xfs_dir2_free_hdr_t) - 1));
+       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC) ||
+              free->hdr.magic == cpu_to_be32(XFS_DIR3_FREE_MAGIC));
+       xfs_trans_log_buf(tp, bp, 0, xfs_dir3_free_hdr_size(tp->t_mountp) - 1);
 }
 
 /*
@@ -183,6 +310,7 @@ xfs_dir2_leaf_to_node(
        xfs_dir2_data_off_t     off;            /* freespace entry value */
        __be16                  *to;            /* pointer to freespace entry */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir3_icfree_hdr freehdr;
 
        trace_xfs_dir2_leaf_to_node(args);
 
@@ -199,43 +327,43 @@ xfs_dir2_leaf_to_node(
        /*
         * Get the buffer for the new freespace block.
         */
-       error = xfs_da_get_buf(tp, dp, xfs_dir2_db_to_da(mp, fdb), -1, &fbp,
-                               XFS_DATA_FORK);
+       error = xfs_dir3_free_get_buf(tp, dp, fdb, &fbp);
        if (error)
                return error;
-       fbp->b_ops = &xfs_dir2_free_buf_ops;
 
        free = fbp->b_addr;
+       xfs_dir3_free_hdr_from_disk(&freehdr, free);
        leaf = lbp->b_addr;
        ltp = xfs_dir2_leaf_tail_p(mp, leaf);
-       /*
-        * Initialize the freespace block header.
-        */
-       free->hdr.magic = cpu_to_be32(XFS_DIR2_FREE_MAGIC);
-       free->hdr.firstdb = 0;
-       ASSERT(be32_to_cpu(ltp->bestcount) <= (uint)dp->i_d.di_size / mp->m_dirblksize);
-       free->hdr.nvalid = ltp->bestcount;
+       ASSERT(be32_to_cpu(ltp->bestcount) <=
+                               (uint)dp->i_d.di_size / mp->m_dirblksize);
+
        /*
         * Copy freespace entries from the leaf block to the new block.
         * Count active entries.
         */
-       for (i = n = 0, from = xfs_dir2_leaf_bests_p(ltp), to = free->bests;
-            i < be32_to_cpu(ltp->bestcount); i++, from++, to++) {
+       from = xfs_dir2_leaf_bests_p(ltp);
+       to = xfs_dir3_free_bests_p(mp, free);
+       for (i = n = 0; i < be32_to_cpu(ltp->bestcount); i++, from++, to++) {
                if ((off = be16_to_cpu(*from)) != NULLDATAOFF)
                        n++;
                *to = cpu_to_be16(off);
        }
-       free->hdr.nused = cpu_to_be32(n);
-
-       lbp->b_ops = &xfs_dir2_leafn_buf_ops;
-       leaf->hdr.info.magic = cpu_to_be16(XFS_DIR2_LEAFN_MAGIC);
 
        /*
-        * Log everything.
+        * Now initialize the freespace block header.
         */
-       xfs_dir2_leaf_log_header(tp, lbp);
+       freehdr.nused = n;
+       freehdr.nvalid = be32_to_cpu(ltp->bestcount);
+
+       xfs_dir3_free_hdr_to_disk(fbp->b_addr, &freehdr);
+       xfs_dir2_free_log_bests(tp, fbp, 0, freehdr.nvalid - 1);
        xfs_dir2_free_log_header(tp, fbp);
-       xfs_dir2_free_log_bests(tp, fbp, 0, be32_to_cpu(free->hdr.nvalid) - 1);
+
+       /* convert the leaf to a leafnode */
+       leaf->hdr.info.magic = cpu_to_be16(XFS_DIR2_LEAFN_MAGIC);
+       lbp->b_ops = &xfs_dir2_leafn_buf_ops;
+       xfs_dir2_leaf_log_header(tp, lbp);
        xfs_dir2_leafn_check(dp, lbp);
        return 0;
 }
@@ -354,6 +482,23 @@ xfs_dir2_leafn_check(
        }
        ASSERT(be16_to_cpu(leaf->hdr.stale) == stale);
 }
+
+static void
+xfs_dir2_free_hdr_check(
+       struct xfs_mount *mp,
+       struct xfs_buf  *bp,
+       xfs_dir2_db_t   db)
+{
+       struct xfs_dir3_icfree_hdr hdr;
+
+       xfs_dir3_free_hdr_from_disk(&hdr, bp->b_addr);
+
+       ASSERT((hdr.firstdb % xfs_dir3_free_max_bests(mp)) == 0);
+       ASSERT(hdr.firstdb <= db);
+       ASSERT(db < hdr.firstdb + hdr.nvalid);
+}
+#else
+#define xfs_dir2_free_hdr_check(mp, dp, db)
 #endif /* DEBUG */
 
 /*
@@ -424,7 +569,8 @@ xfs_dir2_leafn_lookup_for_addname(
                curbp = state->extrablk.bp;
                curfdb = state->extrablk.blkno;
                free = curbp->b_addr;
-               ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
+               ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC) ||
+                      free->hdr.magic == cpu_to_be32(XFS_DIR3_FREE_MAGIC));
        }
        length = xfs_dir2_data_entsize(args->namelen);
        /*
@@ -451,6 +597,8 @@ xfs_dir2_leafn_lookup_for_addname(
                 * in hand, take a look at it.
                 */
                if (newdb != curdb) {
+                       __be16 *bests;
+
                        curdb = newdb;
                        /*
                         * Convert the data block to the free block
@@ -473,13 +621,8 @@ xfs_dir2_leafn_lookup_for_addname(
                                if (error)
                                        return error;
                                free = curbp->b_addr;
-                               ASSERT(be32_to_cpu(free->hdr.magic) ==
-                                       XFS_DIR2_FREE_MAGIC);
-                               ASSERT((be32_to_cpu(free->hdr.firstdb) %
-                                       xfs_dir2_free_max_bests(mp)) == 0);
-                               ASSERT(be32_to_cpu(free->hdr.firstdb) <= curdb);
-                               ASSERT(curdb < be32_to_cpu(free->hdr.firstdb) +
-                                       be32_to_cpu(free->hdr.nvalid));
+
+                               xfs_dir2_free_hdr_check(mp, curbp, curdb);
                        }
                        /*
                         * Get the index for our entry.
@@ -488,8 +631,8 @@ xfs_dir2_leafn_lookup_for_addname(
                        /*
                         * If it has room, return it.
                         */
-                       if (unlikely(free->bests[fi] ==
-                           cpu_to_be16(NULLDATAOFF))) {
+                       bests = xfs_dir3_free_bests_p(mp, free);
+                       if (unlikely(bests[fi] == cpu_to_be16(NULLDATAOFF))) {
                                XFS_ERROR_REPORT("xfs_dir2_leafn_lookup_int",
                                                        XFS_ERRLEVEL_LOW, mp);
                                if (curfdb != newfdb)
@@ -497,7 +640,7 @@ xfs_dir2_leafn_lookup_for_addname(
                                return XFS_ERROR(EFSCORRUPTED);
                        }
                        curfdb = newfdb;
-                       if (be16_to_cpu(free->bests[fi]) >= length)
+                       if (be16_to_cpu(bests[fi]) >= length)
                                goto out;
                }
        }
@@ -511,6 +654,12 @@ out:
                state->extrablk.bp = curbp;
                state->extrablk.index = fi;
                state->extrablk.blkno = curfdb;
+
+               /*
+                * Important: this magic number is not in the buffer - it's for
+                * buffer type information and therefore only the free/data type
+                * matters here, not whether CRCs are enabled or not.
+                */
                state->extrablk.magic = XFS_DIR2_FREE_MAGIC;
        } else {
                state->extravalid = 0;
@@ -898,7 +1047,7 @@ xfs_dir2_leafn_rebalance(
 }
 
 static int
-xfs_dir2_data_block_free(
+xfs_dir3_data_block_free(
        xfs_da_args_t           *args,
        struct xfs_dir2_data_hdr *hdr,
        struct xfs_dir2_free    *free,
@@ -909,57 +1058,66 @@ xfs_dir2_data_block_free(
 {
        struct xfs_trans        *tp = args->trans;
        int                     logfree = 0;
+       __be16                  *bests;
+       struct xfs_dir3_icfree_hdr freehdr;
 
-       if (!hdr) {
-               /* One less used entry in the free table.  */
-               be32_add_cpu(&free->hdr.nused, -1);
-               xfs_dir2_free_log_header(tp, fbp);
+       xfs_dir3_free_hdr_from_disk(&freehdr, free);
 
+       bests = xfs_dir3_free_bests_p(tp->t_mountp, free);
+       if (hdr) {
                /*
-                * If this was the last entry in the table, we can trim the
-                * table size back.  There might be other entries at the end
-                * referring to non-existent data blocks, get those too.
+                * Data block is not empty, just set the free entry to the new
+                * value.
                 */
-               if (findex == be32_to_cpu(free->hdr.nvalid) - 1) {
-                       int     i;              /* free entry index */
+               bests[findex] = cpu_to_be16(longest);
+               xfs_dir2_free_log_bests(tp, fbp, findex, findex);
+               return 0;
+       }
 
-                       for (i = findex - 1; i >= 0; i--) {
-                               if (free->bests[i] != cpu_to_be16(NULLDATAOFF))
-                                       break;
-                       }
-                       free->hdr.nvalid = cpu_to_be32(i + 1);
-                       logfree = 0;
-               } else {
-                       /* Not the last entry, just punch it out.  */
-                       free->bests[findex] = cpu_to_be16(NULLDATAOFF);
-                       logfree = 1;
-               }
-               /*
-                * If there are no useful entries left in the block,
-                * get rid of the block if we can.
-                */
-               if (!free->hdr.nused) {
-                       int error;
+       /* One less used entry in the free table. */
+       freehdr.nused--;
 
-                       error = xfs_dir2_shrink_inode(args, fdb, fbp);
-                       if (error == 0) {
-                               fbp = NULL;
-                               logfree = 0;
-                       } else if (error != ENOSPC || args->total != 0)
-                               return error;
-                       /*
-                        * It's possible to get ENOSPC if there is no
-                        * space reservation.  In this case some one
-                        * else will eventually get rid of this block.
-                        */
+       /*
+        * If this was the last entry in the table, we can trim the table size
+        * back.  There might be other entries at the end referring to
+        * non-existent data blocks, get those too.
+        */
+       if (findex == freehdr.nvalid - 1) {
+               int     i;              /* free entry index */
+
+               for (i = findex - 1; i >= 0; i--) {
+                       if (bests[i] != cpu_to_be16(NULLDATAOFF))
+                               break;
                }
+               freehdr.nvalid = i + 1;
+               logfree = 0;
        } else {
+               /* Not the last entry, just punch it out.  */
+               bests[findex] = cpu_to_be16(NULLDATAOFF);
+               logfree = 1;
+       }
+
+       xfs_dir3_free_hdr_to_disk(free, &freehdr);
+       xfs_dir2_free_log_header(tp, fbp);
+
+       /*
+        * If there are no useful entries left in the block, get rid of the
+        * block if we can.
+        */
+       if (!freehdr.nused) {
+               int error;
+
+               error = xfs_dir2_shrink_inode(args, fdb, fbp);
+               if (error == 0) {
+                       fbp = NULL;
+                       logfree = 0;
+               } else if (error != ENOSPC || args->total != 0)
+                       return error;
                /*
-                * Data block is not empty, just set the free entry to the new
-                * value.
+                * It's possible to get ENOSPC if there is no
+                * space reservation.  In this case some one
+                * else will eventually get rid of this block.
                 */
-               free->bests[findex] = cpu_to_be16(longest);
-               logfree = 1;
        }
 
        /* Log the free entry that changed, unless we got rid of it.  */
@@ -1062,10 +1220,14 @@ xfs_dir2_leafn_remove(
                if (error)
                        return error;
                free = fbp->b_addr;
-               ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
-               ASSERT(be32_to_cpu(free->hdr.firstdb) ==
-                      xfs_dir2_free_max_bests(mp) *
-                      (fdb - XFS_DIR2_FREE_FIRSTDB(mp)));
+#ifdef DEBUG
+       {
+               struct xfs_dir3_icfree_hdr freehdr;
+               xfs_dir3_free_hdr_from_disk(&freehdr, free);
+               ASSERT(freehdr.firstdb == xfs_dir3_free_max_bests(mp) *
+                                         (fdb - XFS_DIR2_FREE_FIRSTDB(mp)));
+       }
+#endif
                /*
                 * Calculate which entry we need to fix.
                 */
@@ -1096,7 +1258,7 @@ xfs_dir2_leafn_remove(
                 * If we got rid of the data block, we can eliminate that entry
                 * in the free block.
                 */
-               error = xfs_dir2_data_block_free(args, hdr, free,
+               error = xfs_dir3_data_block_free(args, hdr, free,
                                                 fdb, findex, fbp, longest);
                if (error)
                        return error;
@@ -1447,6 +1609,8 @@ xfs_dir2_node_addname_int(
        int                     needscan;       /* need to rescan data frees */
        __be16                  *tagp;          /* data entry tag pointer */
        xfs_trans_t             *tp;            /* transaction pointer */
+       __be16                  *bests;
+       struct xfs_dir3_icfree_hdr freehdr;
 
        dp = args->dp;
        mp = dp->i_mount;
@@ -1464,36 +1628,37 @@ xfs_dir2_node_addname_int(
                 */
                ifbno = fblk->blkno;
                free = fbp->b_addr;
-               ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
                findex = fblk->index;
+               bests = xfs_dir3_free_bests_p(mp, free);
+               xfs_dir3_free_hdr_from_disk(&freehdr, free);
+
                /*
                 * This means the free entry showed that the data block had
                 * space for our entry, so we remembered it.
                 * Use that data block.
                 */
                if (findex >= 0) {
-                       ASSERT(findex < be32_to_cpu(free->hdr.nvalid));
-                       ASSERT(be16_to_cpu(free->bests[findex]) != NULLDATAOFF);
-                       ASSERT(be16_to_cpu(free->bests[findex]) >= length);
-                       dbno = be32_to_cpu(free->hdr.firstdb) + findex;
-               }
-               /*
-                * The data block looked at didn't have enough room.
-                * We'll start at the beginning of the freespace entries.
-                */
-               else {
+                       ASSERT(findex < freehdr.nvalid);
+                       ASSERT(be16_to_cpu(bests[findex]) != NULLDATAOFF);
+                       ASSERT(be16_to_cpu(bests[findex]) >= length);
+                       dbno = freehdr.firstdb + findex;
+               } else {
+                       /*
+                        * The data block looked at didn't have enough room.
+                        * We'll start at the beginning of the freespace entries.
+                        */
                        dbno = -1;
                        findex = 0;
                }
-       }
-       /*
-        * Didn't come in with a freespace block, so don't have a data block.
-        */
-       else {
+       } else {
+               /*
+                * Didn't come in with a freespace block, so no data block.
+                */
                ifbno = dbno = -1;
                fbp = NULL;
                findex = 0;
        }
+
        /*
         * If we don't have a data block yet, we're going to scan the
         * freespace blocks looking for one.  Figure out what the
@@ -1547,20 +1712,26 @@ xfs_dir2_node_addname_int(
                        if (!fbp)
                                continue;
                        free = fbp->b_addr;
-                       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
                        findex = 0;
                }
                /*
                 * Look at the current free entry.  Is it good enough?
+                *
+                * The bests initialisation should be where the bufer is read in
+                * the above branch. But gcc is too stupid to realise that bests
+                * and the freehdr are actually initialised if they are placed
+                * there, so we have to do it here to avoid warnings. Blech.
                 */
-               if (be16_to_cpu(free->bests[findex]) != NULLDATAOFF &&
-                   be16_to_cpu(free->bests[findex]) >= length)
-                       dbno = be32_to_cpu(free->hdr.firstdb) + findex;
+               bests = xfs_dir3_free_bests_p(mp, free);
+               xfs_dir3_free_hdr_from_disk(&freehdr, free);
+               if (be16_to_cpu(bests[findex]) != NULLDATAOFF &&
+                   be16_to_cpu(bests[findex]) >= length)
+                       dbno = freehdr.firstdb + findex;
                else {
                        /*
                         * Are we done with the freeblock?
                         */
-                       if (++findex == be32_to_cpu(free->hdr.nvalid)) {
+                       if (++findex == freehdr.nvalid) {
                                /*
                                 * Drop the block.
                                 */
@@ -1614,11 +1785,11 @@ xfs_dir2_node_addname_int(
                 * If there wasn't a freespace block, the read will
                 * return a NULL fbp.  Allocate and initialize a new one.
                 */
-               if( fbp == NULL ) {
-                       if ((error = xfs_dir2_grow_inode(args, XFS_DIR2_FREE_SPACE,
-                                                       &fbno))) {
+               if (!fbp) {
+                       error = xfs_dir2_grow_inode(args, XFS_DIR2_FREE_SPACE,
+                                                   &fbno);
+                       if (error)
                                return error;
-                       }
 
                        if (unlikely(xfs_dir2_db_to_fdb(mp, dbno) != fbno)) {
                                xfs_alert(mp,
@@ -1646,27 +1817,24 @@ xfs_dir2_node_addname_int(
                        /*
                         * Get a buffer for the new block.
                         */
-                       error = xfs_da_get_buf(tp, dp,
-                                              xfs_dir2_db_to_da(mp, fbno),
-                                              -1, &fbp, XFS_DATA_FORK);
+                       error = xfs_dir3_free_get_buf(tp, dp, fbno, &fbp);
                        if (error)
                                return error;
-                       fbp->b_ops = &xfs_dir2_free_buf_ops;
+                       free = fbp->b_addr;
+                       bests = xfs_dir3_free_bests_p(mp, free);
+                       xfs_dir3_free_hdr_from_disk(&freehdr, free);
 
                        /*
-                        * Initialize the new block to be empty, and remember
-                        * its first slot as our empty slot.
+                        * Remember the first slot as our empty slot.
                         */
-                       free = fbp->b_addr;
-                       free->hdr.magic = cpu_to_be32(XFS_DIR2_FREE_MAGIC);
-                       free->hdr.firstdb = cpu_to_be32(
-                               (fbno - XFS_DIR2_FREE_FIRSTDB(mp)) *
-                               xfs_dir2_free_max_bests(mp));
+                       freehdr.firstdb = (fbno - XFS_DIR2_FREE_FIRSTDB(mp)) *
+                                       xfs_dir3_free_max_bests(mp);
                        free->hdr.nvalid = 0;
                        free->hdr.nused = 0;
                } else {
                        free = fbp->b_addr;
-                       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
+                       bests = xfs_dir3_free_bests_p(mp, free);
+                       xfs_dir3_free_hdr_from_disk(&freehdr, free);
                }
 
                /*
@@ -1677,20 +1845,21 @@ xfs_dir2_node_addname_int(
                 * If it's after the end of the current entries in the
                 * freespace block, extend that table.
                 */
-               if (findex >= be32_to_cpu(free->hdr.nvalid)) {
-                       ASSERT(findex < xfs_dir2_free_max_bests(mp));
-                       free->hdr.nvalid = cpu_to_be32(findex + 1);
+               if (findex >= freehdr.nvalid) {
+                       ASSERT(findex < xfs_dir3_free_max_bests(mp));
+                       freehdr.nvalid = findex + 1;
                        /*
                         * Tag new entry so nused will go up.
                         */
-                       free->bests[findex] = cpu_to_be16(NULLDATAOFF);
+                       bests[findex] = cpu_to_be16(NULLDATAOFF);
                }
                /*
                 * If this entry was for an empty data block
                 * (this should always be true) then update the header.
                 */
-               if (free->bests[findex] == cpu_to_be16(NULLDATAOFF)) {
-                       be32_add_cpu(&free->hdr.nused, 1);
+               if (bests[findex] == cpu_to_be16(NULLDATAOFF)) {
+                       freehdr.nused++;
+                       xfs_dir3_free_hdr_to_disk(fbp->b_addr, &freehdr);
                        xfs_dir2_free_log_header(tp, fbp);
                }
                /*
@@ -1699,7 +1868,7 @@ xfs_dir2_node_addname_int(
                 * change again.
                 */
                hdr = dbp->b_addr;
-               free->bests[findex] = hdr->bestfree[0].length;
+               bests[findex] = hdr->bestfree[0].length;
                logfree = 1;
        }
        /*
@@ -1758,8 +1927,9 @@ xfs_dir2_node_addname_int(
        /*
         * If the freespace entry is now wrong, update it.
         */
-       if (be16_to_cpu(free->bests[findex]) != be16_to_cpu(hdr->bestfree[0].length)) {
-               free->bests[findex] = hdr->bestfree[0].length;
+       bests = xfs_dir3_free_bests_p(mp, free); /* gcc is so stupid */
+       if (be16_to_cpu(bests[findex]) != be16_to_cpu(hdr->bestfree[0].length)) {
+               bests[findex] = hdr->bestfree[0].length;
                logfree = 1;
        }
        /*
@@ -1995,6 +2165,7 @@ xfs_dir2_node_trim_free(
        xfs_dir2_free_t         *free;          /* freespace structure */
        xfs_mount_t             *mp;            /* filesystem mount point */
        xfs_trans_t             *tp;            /* transaction pointer */
+       struct xfs_dir3_icfree_hdr freehdr;
 
        dp = args->dp;
        mp = dp->i_mount;
@@ -2012,11 +2183,12 @@ xfs_dir2_node_trim_free(
        if (!bp)
                return 0;
        free = bp->b_addr;
-       ASSERT(free->hdr.magic == cpu_to_be32(XFS_DIR2_FREE_MAGIC));
+       xfs_dir3_free_hdr_from_disk(&freehdr, free);
+
        /*
         * If there are used entries, there's nothing to do.
         */
-       if (be32_to_cpu(free->hdr.nused) > 0) {
+       if (freehdr.nused > 0) {
                xfs_trans_brelse(tp, bp);
                *rvalp = 0;
                return 0;