xfs: fix bugs uncovered by dbench testing

index 31e4ea2..ef8e7d9 100644
@@ -377,9 +377,23 @@ xfs_log_commit_cil(
 	xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
 	xfs_trans_unreserve_and_mod_sb(tp);
 
+	/*
+	 * Once all the items of the transaction have been copied to the CIL,
+	 * the items can be unlocked and freed.
+	 *
+	 * This needs to be done before we drop the CIL context lock because we
+	 * have to update state in the log items and unlock them before they go
+	 * to disk. If we don't, then the CIL checkpoint can race with us and
+	 * we can run checkpoint completion before we've updated and unlocked
+	 * the log items. This affects (at least) processing of stale buffers,
+	 * inodes and EFIs.
+	 */
+	xfs_trans_free_items(tp, *commit_lsn, 0);
+
 	/* check for background commit before unlock */
 	if (log->l_cilp->xc_ctx->space_used > XLOG_CIL_SPACE_LIMIT(log))
 		push = 1;
+
 	up_read(&log->l_cilp->xc_ctx_lock);
 
 	/*

index fdca741..1c47eda 100644
@@ -1167,7 +1167,7 @@ xfs_trans_del_item(
  * Unlock all of the items of a transaction and free all the descriptors
  * of that transaction.
  */
-STATIC void
+void
 xfs_trans_free_items(
 	struct xfs_trans	*tp,
 	xfs_lsn_t		commit_lsn,

@@ -1653,9 +1653,6 @@ xfs_trans_commit_cil(
 		return error;
 
 	current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
-
-	/* xfs_trans_free_items() unlocks them first */
-	xfs_trans_free_items(tp, *commit_lsn, 0);
 	xfs_trans_free(tp);
 	return 0;
 }

index e2d93d8..62da86c 100644
@@ -25,7 +25,8 @@ struct xfs_trans;
 
 void	xfs_trans_add_item(struct xfs_trans *, struct xfs_log_item *);
 void	xfs_trans_del_item(struct xfs_log_item *);
-
+void	xfs_trans_free_items(struct xfs_trans *tp, xfs_lsn_t commit_lsn,
+				int flags);
 void	xfs_trans_item_committed(struct xfs_log_item *lip,
 				xfs_lsn_t commit_lsn, int aborted);
 void	xfs_trans_unreserve_and_mod_sb(struct xfs_trans *tp);

index 68415cb..abc50b6 100644
@@ -1914,6 +1914,11 @@ xfs_iunlink_remove(
 	return 0;
 }
 
+/*
+ * A big issue when freeing the inode cluster is is that we _cannot_ skip any
+ * inodes that are in memory - they all must be marked stale and attached to
+ * the cluster buffer.
+ */
 STATIC void
 xfs_ifree_cluster(
 	xfs_inode_t	*free_ip,

@@ -1945,8 +1950,6 @@ xfs_ifree_cluster(
 	}
 
 	for (j = 0; j < nbufs; j++, inum += ninodes) {
-		int	found = 0;
-
 		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
 					 XFS_INO_TO_AGBNO(mp, inum));
 

@@ -1963,26 +1966,6 @@ xfs_ifree_cluster(
 					XBF_LOCK);
 
 		/*
-		 * Walk the inodes already attached to the buffer and mark them
-		 * stale. These will all have the flush locks held, so an
-		 * in-memory inode walk can't lock them.
-		 */
-		lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
-		while (lip) {
-			if (lip->li_type == XFS_LI_INODE) {
-				iip = (xfs_inode_log_item_t *)lip;
-				ASSERT(iip->ili_logged == 1);
-				lip->li_cb = xfs_istale_done;
-				xfs_trans_ail_copy_lsn(mp->m_ail,
-							&iip->ili_flush_lsn,
-							&iip->ili_item.li_lsn);
-				xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
-				found++;
-			}
-			lip = lip->li_bio_list;
-		}
-
-		/*
 		 * For each inode in memory attempt to add it to the inode
 		 * buffer and set it up for being staled on buffer IO
 		 * completion.  This is safe as we've locked out tail pushing

@@ -1999,42 +1982,68 @@ xfs_ifree_cluster(
 
 			/* Inode not in memory or stale, nothing to do */
 			if (!ip || xfs_iflags_test(ip, XFS_ISTALE)) {
+next_inode:
 				read_unlock(&pag->pag_ici_lock);
 				continue;
 			}
 
-			/* don't try to lock/unlock the current inode */
+			/*
+			 * Walk the inodes already attached to the buffer to
+			 * see if this inode is already there. If so, mark it
+			 * stale. These will all have the flush locks held, so an
+			 * in-memory inode walk can't lock them. We do this
+			 * inside the loop so that we know for certain that we
+			 * don't encounter inodes with the flush lock already
+			 * held and hence deadlock below.
+			 */
+			lip = XFS_BUF_FSPRIVATE(bp, xfs_log_item_t *);
+			do {
+				if (lip->li_type == XFS_LI_INODE) {
+					iip = (xfs_inode_log_item_t *)lip;
+					if (iip->ili_inode != ip)
+						continue;
+					ASSERT(iip->ili_logged == 1);
+					lip->li_cb = xfs_istale_done;
+					xfs_trans_ail_copy_lsn(mp->m_ail,
+								&iip->ili_flush_lsn,
+								&iip->ili_item.li_lsn);
+					xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
+					goto next_inode;
+				}
+			} while ((lip = lip->li_bio_list) != NULL);
+
+			/*
+			 * Don't try to lock/unlock the current inode, but we
+			 * _cannot_ skip the other inodes that we did not find
+			 * in the list attached to the buffer and are not
+			 * already marked stale. If we can't lock it, back off
+			 * and retry.
+			 */
+
 			if (ip != free_ip &&
 			    !xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
 				read_unlock(&pag->pag_ici_lock);
+				delay(1);
+				i--;
 				continue;
 			}
 			read_unlock(&pag->pag_ici_lock);
 
-			if (!xfs_iflock_nowait(ip)) {
-				if (ip != free_ip)
-					xfs_iunlock(ip, XFS_ILOCK_EXCL);
-				continue;
-			}
-
+			xfs_iflock(ip);
 			xfs_iflags_set(ip, XFS_ISTALE);
-			if (xfs_inode_clean(ip)) {
-				ASSERT(ip != free_ip);
-				xfs_ifunlock(ip);
-				xfs_iunlock(ip, XFS_ILOCK_EXCL);
-				continue;
-			}
 
+			/*
+			 * we don't need to attach clean inodes or those only
+			 * with unlogged changes (which we throw away, anyway).
+			 */
 			iip = ip->i_itemp;
-			if (!iip) {
-				/* inode with unlogged changes only */
+			if (!iip || xfs_inode_clean(ip)) {
 				ASSERT(ip != free_ip);
 				ip->i_update_core = 0;
 				xfs_ifunlock(ip);
 				xfs_iunlock(ip, XFS_ILOCK_EXCL);
 				continue;
 			}
-			found++;
 
 			iip->ili_last_fields = iip->ili_format.ilf_fields;
 			iip->ili_format.ilf_fields = 0;

@@ -2049,8 +2058,7 @@ xfs_ifree_cluster(
 				xfs_iunlock(ip, XFS_ILOCK_EXCL);
 		}
 
-		if (found)
-			xfs_trans_stale_inode_buf(tp, bp);
+		xfs_trans_stale_inode_buf(tp, bp);
 		xfs_trans_binval(tp, bp);
 	}