journal_clean_checkpoint_list() is supposed to release any written-back
checkpointed buffers from the journalling system.  The number of such buffers
is potentially enough to cover the whole journal - that's a lot with a
maximum-sized journal of 409600 blocks.

The function is fairly lame, really.  A better solution would be to hook into
the blockdev mapping's ->releasepage a_op, but that's messy.

Change it so that it will drop locks and schedule away if it is holding off
preemption.  Rotate the where-to-start-next-time pointers to try to avoid the
obvious quadratic search problems.

Note that it can still hold off preemption for a long time if lots of those
buffers are dirty.  We cannot do much about that if livelock is to be avoided.

Signed-off-by: Andrew Morton <akpm@osdl.org>
Index: linux-2.6.8-rc1/fs/jbd/checkpoint.c
===================================================================
--- linux-2.6.8-rc1.orig/fs/jbd/checkpoint.c	2004-07-15 10:03:27.623032107 +1000
+++ linux-2.6.8-rc1/fs/jbd/checkpoint.c	2004-07-15 10:03:35.040851526 +1000
@@ -462,9 +462,8 @@
  *
  * Find all the written-back checkpoint buffers in the journal and release them.
  *
- * Called with the journal locked.
- * Called with j_list_lock held.
- * Returns number of bufers reaped (for debug)
+ * Called with j_list_lock held, drops it.
+ * Returns number of bufers reaped
  */
 
 int __journal_clean_checkpoint_list(journal_t *journal)
@@ -474,7 +473,7 @@
 
 	transaction = journal->j_checkpoint_transactions;
 	if (transaction == 0)
-		goto out;
+		goto out_unlock;
 
 	last_transaction = transaction->t_cpprev;
 	next_transaction = transaction;
@@ -491,13 +490,41 @@
 			do {
 				jh = next_jh;
 				next_jh = jh->b_cpnext;
-				/* Use trylock because of the ranknig */
+				/* Use trylock because of the ranking */
 				if (jbd_trylock_bh_state(jh2bh(jh)))
 					ret += __try_to_free_cp_buf(jh);
 			} while (jh != last_jh);
 		}
+#ifdef CONFIG_PREEMPT
+		/*
+		 * This is potentially sucky: semi-quadratic performance if
+		 * there are a lot of dirty buffers.  So only do it if the user
+		 * has chosen a preemptible kernel.  If !CONFIG_PREEMPT we're
+		 * optimimising for straight-line performance, after all.
+		 * We don't test cond_resched() here because another CPU could
+		 * be waiting on j_list_lock() while holding a different lock.
+		 */
+		if ((ret & 127) == 127) {
+			spin_unlock(&journal->j_list_lock);
+			/*
+			 * We need to schedule away.  Rotate both this
+			 * transaction's buffer list and the checkpoint list to
+			 * try to avoid quadratic behaviour.
+			 */
+			jh = transaction->t_checkpoint_list;
+			if (jh)
+				transaction->t_checkpoint_list = jh->b_cpnext;
+
+			transaction = journal->j_checkpoint_transactions;
+			if (transaction)
+				journal->j_checkpoint_transactions =
+					transaction->t_cpnext;
+			return ret;
+		}
+#endif
 	} while (transaction != last_transaction);
-out:
+out_unlock:
+	spin_unlock(&journal->j_list_lock);
 	return ret;
 }
 
Index: linux-2.6.8-rc1/fs/jbd/commit.c
===================================================================
--- linux-2.6.8-rc1.orig/fs/jbd/commit.c	2004-07-15 02:41:40.000000000 +1000
+++ linux-2.6.8-rc1/fs/jbd/commit.c	2004-07-15 10:03:35.057848821 +1000
@@ -208,9 +208,16 @@
 	 * checkpoint lists.  We do this *before* commit because it potentially
 	 * frees some memory
 	 */
-	spin_lock(&journal->j_list_lock);
-	__journal_clean_checkpoint_list(journal);
-	spin_unlock(&journal->j_list_lock);
+	spin_unlock(&journal->j_state_lock);
+	{
+		int nr_cleaned;
+
+		do {
+			spin_lock(&journal->j_list_lock);
+			nr_cleaned = __journal_clean_checkpoint_list(journal);
+		} while (nr_cleaned);
+	}
+	spin_lock(&journal->j_state_lock);
 
 	jbd_debug (3, "JBD: commit phase 1\n");