From: NeilBrown <neilb@suse.de>
Date: Mon, 21 Jun 2010 00:39:00 +0000 (+1000)
Subject: Don't let writepage spoil a transaction.
X-Git-Url: http://git.neil.brown.name/?a=commitdiff_plain;h=46b138281062b314cf5a6ca7d4e2416bece7de27;p=LaFS.git

Don't let writepage spoil a transaction.

If a block is involved in a transaction (e.g. dir update) we
mustn't allow writepage to flush the page until the transaction
completes.

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

diff --git a/file.c b/file.c
index 6fb1732..69b6a88 100644
--- a/file.c
+++ b/file.c
@@ -218,8 +218,15 @@ lafs_writepage(struct page *page, struct writeback_control *wbc)
 				rv = AOP_WRITEPAGE_ACTIVATE;
 			else {
 				lafs_iolock_written(&b->b);
-				/* block might have been invalidated while we waited */
-				if (test_bit(B_Dirty, &b->b.flags))
+				/* block might have been invalidated,
+				 * or Pinned, while we waited */
+				/* FIXME do I need to test if the iblock has appeared
+				 * while we waited too??
+				 */
+				if (test_bit(B_PinPending, &b->b.flags)) {
+					rv = AOP_WRITEPAGE_ACTIVATE;
+					lafs_iounlock_block(&b->b);
+				} else if (test_bit(B_Dirty, &b->b.flags))
 					lafs_cluster_allocate(&b->b, 0);
 				else
 					lafs_iounlock_block(&b->b);