Cleaner: add a memory barrier to ensure we see i_size promptly.

There is a possible race that we need a barrier to protect against.
I think... or hope.

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

diff --git a/clean.c b/clean.c
index bb7a3c26c75e18e366285e3225f1fe7110e8db39..6cc09f5086440326752840c21dd17f2e427e0921 100644 (file)
--- a/clean.c
+++ b/clean.c
@@ -340,8 +340,17 @@ static void cleaner_parse(struct fs *fs, struct toclean *tc)
                                        list_add_tail(&b->cleaning, &tc->cleaning);
                        }
 
-                       /* FIXME do I need a memory barrier. to ensure truncate
-                        * sees the not-list_empty, and we see i_size? */
+                       /* We can race with truncate here, so need to check
+                        * i_size again now that b->cleaning is non-empty.
+                        * The thread doing the truncate will have to lock
+                        * the page holding this block, which should be enough
+                        * of a barrier so that if it sets i_size after now,
+                        * it will see that b->cleaning is non-empty.
+                        * We need to be sure that if it sets it before now,
+                        * we get to see i_size.
+                        * So I think a memory barrier is a good idea...
+                        */
+                       mb();
 
                        if (LAFSI(ino)->type == 0 ||
                            (LAFSI(ino)->type >= TypeBase &&