}
 
 /*
- * Move half num of the xattrs in old bucket(blk) to new bucket(new_blk).
+ * Find the suitable pos when we divide a bucket into 2.
+ * We have to make sure the xattrs with the same hash value exist
+ * in the same bucket.
+ *
+ * If this ocfs2_xattr_header covers more than one hash value, find a
+ * place where the hash value changes.  Try to find the most even split.
+ * The most common case is that all entries have different hash values,
+ * and the first check we make will find a place to split.
+ */
+static int ocfs2_xattr_find_divide_pos(struct ocfs2_xattr_header *xh)
+{
+       struct ocfs2_xattr_entry *entries = xh->xh_entries;
+       int count = le16_to_cpu(xh->xh_count);
+       int delta, middle = count / 2;
+
+       /*
+        * We start at the middle.  Each step gets farther away in both
+        * directions.  We therefore hit the change in hash value
+        * nearest to the middle.  Note that this loop does not execute for
+        * count < 2.
+        */
+       for (delta = 0; delta < middle; delta++) {
+               /* Let's check delta earlier than middle */
+               if (cmp_xe(&entries[middle - delta - 1],
+                          &entries[middle - delta]))
+                       return middle - delta;
+
+               /* For even counts, don't walk off the end */
+               if ((middle + delta + 1) == count)
+                       continue;
+
+               /* Now try delta past middle */
+               if (cmp_xe(&entries[middle + delta],
+                          &entries[middle + delta + 1]))
+                       return middle + delta + 1;
+       }
+
+       /* Every entry had the same hash */
+       return count;
+}
+
+/*
+ * Move some xattrs in old bucket(blk) to new bucket(new_blk).
  * first_hash will record the 1st hash of the new bucket.
+ *
+ * Normally half of the xattrs will be moved.  But we have to make
+ * sure that the xattrs with the same hash value are stored in the
+ * same bucket. If all the xattrs in this bucket have the same hash
+ * value, the new bucket will be initialized as an empty one and the
+ * first_hash will be initialized as (hash_value+1).
  */
-static int ocfs2_half_xattr_bucket(struct inode *inode,
-                                  handle_t *handle,
-                                  u64 blk,
-                                  u64 new_blk,
-                                  u32 *first_hash,
-                                  int new_bucket_head)
+static int ocfs2_divide_xattr_bucket(struct inode *inode,
+                                   handle_t *handle,
+                                   u64 blk,
+                                   u64 new_blk,
+                                   u32 *first_hash,
+                                   int new_bucket_head)
 {
        int ret, i;
-       u16 count, start, len, name_value_len, xe_len, name_offset;
+       int count, start, len, name_value_len = 0, xe_len, name_offset = 0;
        u16 blk_per_bucket = ocfs2_blocks_per_xattr_bucket(inode->i_sb);
        struct buffer_head **s_bhs, **t_bhs = NULL;
        struct ocfs2_xattr_header *xh;
        struct ocfs2_xattr_entry *xe;
        int blocksize = inode->i_sb->s_blocksize;
 
-       mlog(0, "move half of xattrs from bucket %llu to %llu\n",
+       mlog(0, "move some of xattrs from bucket %llu to %llu\n",
             blk, new_blk);
 
        s_bhs = kcalloc(blk_per_bucket, sizeof(struct buffer_head *), GFP_NOFS);
                }
        }
 
+       xh = (struct ocfs2_xattr_header *)s_bhs[0]->b_data;
+       count = le16_to_cpu(xh->xh_count);
+       start = ocfs2_xattr_find_divide_pos(xh);
+
+       if (start == count) {
+               xe = &xh->xh_entries[start-1];
+
+               /*
+                * initialized a new empty bucket here.
+                * The hash value is set as one larger than
+                * that of the last entry in the previous bucket.
+                */
+               for (i = 0; i < blk_per_bucket; i++)
+                       memset(t_bhs[i]->b_data, 0, blocksize);
+
+               xh = (struct ocfs2_xattr_header *)t_bhs[0]->b_data;
+               xh->xh_free_start = cpu_to_le16(blocksize);
+               xh->xh_entries[0].xe_name_hash = xe->xe_name_hash;
+               le32_add_cpu(&xh->xh_entries[0].xe_name_hash, 1);
+
+               goto set_num_buckets;
+       }
+
        /* copy the whole bucket to the new first. */
        for (i = 0; i < blk_per_bucket; i++)
                memcpy(t_bhs[i]->b_data, s_bhs[i]->b_data, blocksize);
 
        /* update the new bucket. */
        xh = (struct ocfs2_xattr_header *)t_bhs[0]->b_data;
-       count = le16_to_cpu(xh->xh_count);
-       start = count / 2;
 
        /*
         * Calculate the total name/value len and xh_free_start for
                        xh->xh_free_start = xe->xe_name_offset;
        }
 
+set_num_buckets:
        /* set xh->xh_num_buckets for the new xh. */
        if (new_bucket_head)
                xh->xh_num_buckets = cpu_to_le16(1);
                *first_hash = le32_to_cpu(xh->xh_entries[0].xe_name_hash);
 
        /*
-        * Now only update the 1st block of the old bucket.
-        * Please note that the entry has been sorted already above.
+        * Now only update the 1st block of the old bucket.  If we
+        * just added a new empty bucket, there is no need to modify
+        * it.
         */
+       if (start == count)
+               goto out;
+
        xh = (struct ocfs2_xattr_header *)s_bhs[0]->b_data;
        memset(&xh->xh_entries[start], 0,
               sizeof(struct ocfs2_xattr_entry) * (count - start));
 }
 
 /*
- * Move half of the xattrs in this cluster to the new cluster.
+ * Move some xattrs in this cluster to the new cluster.
  * This function should only be called when bucket size == cluster size.
  * Otherwise ocfs2_mv_xattr_bucket_cross_cluster should be used instead.
  */
-static int ocfs2_half_xattr_cluster(struct inode *inode,
-                                   handle_t *handle,
-                                   u64 prev_blk,
-                                   u64 new_blk,
-                                   u32 *first_hash)
+static int ocfs2_divide_xattr_cluster(struct inode *inode,
+                                     handle_t *handle,
+                                     u64 prev_blk,
+                                     u64 new_blk,
+                                     u32 *first_hash)
 {
        u16 blk_per_bucket = ocfs2_blocks_per_xattr_bucket(inode->i_sb);
        int ret, credits = 2 * blk_per_bucket;
        }
 
        /* Move half of the xattr in start_blk to the next bucket. */
-       return  ocfs2_half_xattr_bucket(inode, handle, prev_blk,
-                                       new_blk, first_hash, 1);
+       return  ocfs2_divide_xattr_bucket(inode, handle, prev_blk,
+                                         new_blk, first_hash, 1);
 }
 
 /*
                                                     last_blk, new_blk,
                                                     v_start);
                else {
-                       ret = ocfs2_half_xattr_cluster(inode, handle,
-                                                      last_blk, new_blk,
-                                                      v_start);
+                       ret = ocfs2_divide_xattr_cluster(inode, handle,
+                                                        last_blk, new_blk,
+                                                        v_start);
 
                        if ((*header_bh)->b_blocknr == last_blk && extend)
                                *extend = 0;
        }
 
        /* Move half of the xattr in start_blk to the next bucket. */
-       ret = ocfs2_half_xattr_bucket(inode, handle, start_blk,
-                                     start_blk + blk_per_bucket, NULL, 0);
+       ret = ocfs2_divide_xattr_bucket(inode, handle, start_blk,
+                                       start_blk + blk_per_bucket, NULL, 0);
 
        le16_add_cpu(&first_xh->xh_num_buckets, 1);
        ocfs2_journal_dirty(handle, first_bh);
        return ret;
 }
 
-/* check whether the xattr bucket is filled up with the same hash value. */
+/*
+ * check whether the xattr bucket is filled up with the same hash value.
+ * If we want to insert the xattr with the same hash, return -ENOSPC.
+ * If we want to insert a xattr with different hash value, go ahead
+ * and ocfs2_divide_xattr_bucket will handle this.
+ */
 static int ocfs2_check_xattr_bucket_collision(struct inode *inode,
-                                             struct ocfs2_xattr_bucket *bucket)
+                                             struct ocfs2_xattr_bucket *bucket,
+                                             const char *name)
 {
        struct ocfs2_xattr_header *xh = bucket->xh;
+       u32 name_hash = ocfs2_xattr_name_hash(inode, name, strlen(name));
+
+       if (name_hash != le32_to_cpu(xh->xh_entries[0].xe_name_hash))
+               return 0;
 
        if (xh->xh_entries[le16_to_cpu(xh->xh_count) - 1].xe_name_hash ==
            xh->xh_entries[0].xe_name_hash) {
                 * one bucket's worth, so check it here whether we need to
                 * add a new bucket for the insert.
                 */
-               ret = ocfs2_check_xattr_bucket_collision(inode, &xs->bucket);
+               ret = ocfs2_check_xattr_bucket_collision(inode,
+                                                        &xs->bucket,
+                                                        xi->name);
                if (ret) {
                        mlog_errno(ret);
                        goto out;