Re: 2.6.8-rc2-mm1

[Andrew Morton <akpm@osdl.org>]

Zwane Mwaikambo <zwane@linuxpower.ca> wrote:
>
> Ingo i believe you have a patch for this, could you push it to Andrew?

I suspect Ingo's patch will be livelockable under some circumstances.

I suspect mine is too, only less so.

> I reckon it's provoked by CONFIG_PREEMPT.

This should fix.


diff -puN fs/jbd/checkpoint.c~journal_clean_checkpoint_list-latency-fix-fix fs/jbd/checkpoint.c
--- 25/fs/jbd/checkpoint.c~journal_clean_checkpoint_list-latency-fix-fix	2004-07-31 11:43:39.320530424 -0700
+++ 25-akpm/fs/jbd/checkpoint.c	2004-07-31 11:44:11.859583736 -0700
@@ -497,8 +497,7 @@ int __journal_clean_checkpoint_list(jour
 		 * 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);
+		if (transaction && (ret & 127) == 127) {
 			/*
 			 * We need to schedule away.  Rotate both this
 			 * transaction's buffer list and the checkpoint list to
@@ -512,6 +511,7 @@ int __journal_clean_checkpoint_list(jour
 			if (transaction)
 				journal->j_checkpoint_transactions =
 					transaction->t_cpnext;
+			spin_unlock(&journal->j_list_lock);
 			return ret;
 		}
 #endif
_