From: NeilBrown <neilb@suse.de>
Date: Mon, 18 Oct 2010 00:48:22 +0000 (+1100)
Subject: Fix lafs_seg_next to talk write-cluster properly.
X-Git-Url: http://git.neil.brown.name/?a=commitdiff_plain;h=19fd807f2b4e86d2f19ae2f345415e140df350de;p=LaFS.git

Fix lafs_seg_next to talk write-cluster properly.

When we have a 2D layout, we want to walk down columns before
across rows, as that encourages adjacent blocks.

So fix lafs_seg_next to do the right thing.

Also check for exceeding the declared size of the segment and
react accordingly.

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

diff --git a/cluster.c b/cluster.c
index 5e188ce..a76b0f5 100644
--- a/cluster.c
+++ b/cluster.c
@@ -340,6 +340,13 @@ static u64 seg_addr(struct fs *fs, struct segpos *seg)
 		/* Setting 'next' address for last cluster in
 		 * a cleaner segment */
 		return 0;
+	if (seg->table > seg->nxt_table ||
+	    (seg->table == seg->nxt_table &&
+	     seg->row >= seg->nxt_row))
+		/* We have gone beyond the end of the cluster,
+		 * must be bad data during roll-forward
+		 */
+		return 0;
 	addr = seg->col * dv->stride;
 	addr += seg->row;
 	addr += seg->table * dv->rows_per_table;
@@ -352,19 +359,38 @@ u64 lafs_seg_next(struct fs *fs, struct segpos *seg)
 {
 	/* step forward one block, returning the address of
 	 * the block stepped over
+	 * The write cluster can have three sections:
+	 * - the tail end of one table
+	 * - a number of complete tables
+	 * - the head of one table
+	 *
+	 * For each table or partial table we want to talk a full
+	 * column at a time, then go to the next column.
+	 * So we step to the next row.  If it is beyond the range of
+	 * the current table, go to 'start' of next column, or to
+	 * next table.
 	 */
 	struct fs_dev *dv = &fs->devs[seg->dev];
 	u64 addr = seg_addr(fs, seg);
 
+	if (!addr)
+		/* Beyond end of cluster */
+		return 0;
+
 	/* now step forward in column or table or seg */
-	seg->col++;
-	if (seg->col >= dv->width) {
-		seg->col = 0;
-		seg->row++;
-		if (seg->row >= dv->rows_per_table) {
-			seg->row = 0;
+	seg->row++;
+	if (seg->row >= dv->rows_per_table ||
+	    (seg->table == seg->nxt_table
+	     && seg->row >= seg->nxt_row)) {
+		seg->col++;
+		if (seg->col >= dv->width) {
+			seg->col = 0;
 			seg->table++;
 		}
+		if (seg->table == seg->st_table)
+			seg->row = seg->st_row;
+		else
+			seg->row = 0;
 	}
 	return addr;
 }
@@ -1176,6 +1202,7 @@ static void cluster_flush(struct fs *fs, int cnum)
 			current_block++;
 		cluster_incdesc(wc, desc_start, b, fs->blocksize_bits);
 		addr = lafs_seg_next(fs, &wc->seg);
+		BUG_ON(!addr);
 		if (cnum && test_bit(B_Dirty, &b->flags))
 			/* We are cleaning but this block is now dirty.
 			 * Don't waste the UnincCredit on recording the
diff --git a/roll.c b/roll.c
index 1649be3..4a6938d 100644
--- a/roll.c
+++ b/roll.c
@@ -568,6 +568,15 @@ roll_one(struct fs *fs, u64 *addrp, struct page *p, struct page *pg,
 						      * response */
 				/* FIXME range check count */
 				while (!err && cnt--) {
+					if (bytes != DescHole &&
+					    !baddr) {
+						/* We have fallen off the end of
+						 * the write-cluster - something
+						 * is wrong with the header
+						 */
+						printk(KERN_WARNING "LAFS: cluster size is wrong\n");
+						return -EIO;
+					}
 					if (!flg && bytes != DescIndex)
 						err = roll_block(fs, fsnum, inum, trunc,
 								 bnum, baddr,