diff --git a/drivers/md/md.c b/drivers/md/md.c
index 64b35cbf3159..bad324171cad 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -590,14 +590,41 @@ static int sb_equal(mdp_super_t *sb1, mdp_super_t *sb2)
 	return ret;
 }
 
+
+static u32 md_csum_fold(u32 csum)
+{
+	csum = (csum & 0xffff) + (csum >> 16);
+	return (csum & 0xffff) + (csum >> 16);
+}
+
 static unsigned int calc_sb_csum(mdp_super_t * sb)
 {
+	u64 newcsum = 0;
+	u32 *sb32 = (u32*)sb;
+	int i;
 	unsigned int disk_csum, csum;
 
 	disk_csum = sb->sb_csum;
 	sb->sb_csum = 0;
-	csum = csum_partial((void *)sb, MD_SB_BYTES, 0);
+
+	for (i = 0; i < MD_SB_BYTES/4 ; i++)
+		newcsum += sb32[i];
+	csum = (newcsum & 0xffffffff) + (newcsum>>32);
+
+
+#ifdef CONFIG_ALPHA
+	/* This used to use csum_partial, which was wrong for several
+	 * reasons including that different results are returned on
+	 * different architectures.  It isn't critical that we get exactly
+	 * the same return value as before (we always csum_fold before
+	 * testing, and that removes any differences).  However as we
+	 * know that csum_partial always returned a 16bit value on
+	 * alphas, do a fold to maximise conformity to previous behaviour.
+	 */
+	sb->sb_csum = md_csum_fold(disk_csum);
+#else
 	sb->sb_csum = disk_csum;
+#endif
 	return csum;
 }
 
@@ -685,7 +712,7 @@ static int super_90_load(mdk_rdev_t *rdev, mdk_rdev_t *refdev, int minor_version
 	if (sb->raid_disks <= 0)
 		goto abort;
 
-	if (csum_fold(calc_sb_csum(sb)) != csum_fold(sb->sb_csum)) {
+	if (md_csum_fold(calc_sb_csum(sb)) != md_csum_fold(sb->sb_csum)) {
 		printk(KERN_WARNING "md: invalid superblock checksum on %s\n",
 			b);
 		goto abort;