From: NeilBrown <neilb@suse.de>
Date: Mon, 21 Jun 2010 00:11:33 +0000 (+1000)
Subject: Temp fix for delaying youth updates.
X-Git-Url: http://git.neil.brown.name/?a=commitdiff_plain;h=f2802c2f09514e0f7272be9adcf3f1dcdc68bfe7;p=LaFS.git

Temp fix for delaying youth updates.

When we allocate a new segment during checkpoint we need to
delay the youth block update, possibly until the roll-forward.

This is just a simple hack to avoid the worst of the problem
but we need to properly delay it at some stage.

Signed-off-by: NeilBrown <neilb@suse.de>
---

diff --git a/roll.c b/roll.c
index 6978b68..1092eb8 100644
--- a/roll.c
+++ b/roll.c
@@ -506,6 +506,9 @@ static int roll_forward(struct fs *fs)
 	buf = kmalloc(max, GFP_KERNEL);
 	if (buf)
 		while (first != next) {
+			/* FIXME if this cluster is in a new segment,
+			 * we need to set the youth number for it.
+			 */
 			err = roll_one(fs, &first, p, pg, max);
 			if (err)
 				break;
diff --git a/segments.c b/segments.c
index 10375fc..d6f3590 100644
--- a/segments.c
+++ b/segments.c
@@ -1035,6 +1035,17 @@ void lafs_free_get(struct fs *fs, unsigned int *dev, u32 *seg,
 	*dev = ss->dev;
 	*seg = ss->segment;
 
+	if (test_bit(FinalCheckpoint, &fs->fsstate)) {
+		/* skip youth update, roll forward will do it. */
+		/* FIXME should more generally delay the update if we
+		 * are in any checkpoint.
+		 */
+		spin_unlock(&fs->lock);
+		if (db)
+			putdref(db, MKREF(youth));
+		return;
+	}
+
 	if (db &&
 	    db->b.inode == fs->devs[*dev].segsum &&
 	    db->b.fileaddr == ((*seg) >> (fs->prime_sb->s_blocksize_bits