xfs: copy li_lsn before dropping AIL lock

Access to log items on the AIL is generally protected by m_ail_lock;
this is particularly needed when we're getting or setting the 64-bit
li_lsn on a 32-bit platform. This patch fixes a couple places where we
were accessing the log item after dropping the AIL lock on 32-bit
machines.

This can result in a partially-zeroed log->l_tail_lsn if
xfs_trans_ail_delete is racing with xfs_trans_ail_update, and in at
least some cases, this can leave the l_tail_lsn with a zero cycle
number, which means xlog_space_left will think the log is full (unless
CONFIG_XFS_DEBUG is set, in which case we'll trip an ASSERT), leading to
processes stuck forever in xlog_grant_log_space.

Thanks to Adrian VanderSpek for first spotting the race potential and to
Dave Chinner for debug assistance.

Signed-off-by: Nathaniel W. Turner <[e-mail]>
Reviewed-by: Christoph Hellwig <[e-mail]>
Signed-off-by: Alex Elder <[e-mail]>

fs/xfs/xfs_trans_ail.c

···
 {
 	xfs_log_item_t		*dlip = NULL;
 	xfs_log_item_t		*mlip;	/* ptr to minimum lip */
+	xfs_lsn_t		tail_lsn;
 
 	mlip = xfs_ail_min(ailp);
 
···
 
 	if (mlip == dlip) {
 		mlip = xfs_ail_min(ailp);
+		/*
+		 * It is not safe to access mlip after the AIL lock is
+		 * dropped, so we must get a copy of li_lsn before we do
+		 * so.  This is especially important on 32-bit platforms
+		 * where accessing and updating 64-bit values like li_lsn
+		 * is not atomic.
+		 */
+		tail_lsn = mlip->li_lsn;
 		spin_unlock(&ailp->xa_lock);
-		xfs_log_move_tail(ailp->xa_mount, mlip->li_lsn);
+		xfs_log_move_tail(ailp->xa_mount, tail_lsn);
 	} else {
 		spin_unlock(&ailp->xa_lock);
 	}
···
 {
 	xfs_log_item_t		*dlip;
 	xfs_log_item_t		*mlip;
+	xfs_lsn_t		tail_lsn;
 
 	if (lip->li_flags & XFS_LI_IN_AIL) {
 		mlip = xfs_ail_min(ailp);
···
 
 		if (mlip == dlip) {
 			mlip = xfs_ail_min(ailp);
+			/*
+			 * It is not safe to access mlip after the AIL lock
+			 * is dropped, so we must get a copy of li_lsn
+			 * before we do so.  This is especially important
+			 * on 32-bit platforms where accessing and updating
+			 * 64-bit values like li_lsn is not atomic.
+			 */
+			tail_lsn = mlip ? mlip->li_lsn : 0;
 			spin_unlock(&ailp->xa_lock);
-			xfs_log_move_tail(ailp->xa_mount,
-						(mlip ? mlip->li_lsn : 0));
+			xfs_log_move_tail(ailp->xa_mount, tail_lsn);
 		} else {
 			spin_unlock(&ailp->xa_lock);
 		}