From 6ddd119256b149f49118a8c29ef7d3d82c7e9f0b Mon Sep 17 00:00:00 2001
From: jonathan <jonathan@NetBSD.org>
Date: Mon, 17 Nov 2003 00:28:32 +0000
Subject: [PATCH] Commit fix for NFS write deadlock, on filesystems mounted via
 local-loopback (lo0). As posted for review on tech-kern 2003-18-09, with a
 long  comment explaining (one of) the deadlock scenarios.

I've used this since shortly after 2002-09-12-, without noticing
performance degradataion or instability for non-loopback mounts.
---
 sys/nfs/nfs_bio.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/sys/nfs/nfs_bio.c b/sys/nfs/nfs_bio.c
index 1a534f0d6ca5..590e69c5cf06 100644
--- a/sys/nfs/nfs_bio.c
+++ b/sys/nfs/nfs_bio.c
@@ -1,4 +1,4 @@
-/*	$NetBSD: nfs_bio.c,v 1.110 2003/09/26 11:51:53 yamt Exp $	*/
+/*	$NetBSD: nfs_bio.c,v 1.111 2003/11/17 00:28:32 jonathan Exp $	*/
 
 /*
  * Copyright (c) 1989, 1993
@@ -35,7 +35,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: nfs_bio.c,v 1.110 2003/09/26 11:51:53 yamt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: nfs_bio.c,v 1.111 2003/11/17 00:28:32 jonathan Exp $");
 
 #include "opt_nfs.h"
 #include "opt_ddb.h"
@@ -836,10 +836,16 @@ again:
 
 	/*
 	 * If we have an iod which can process the request, then queue
-	 * the buffer.
+	 * the buffer.  However, even if we have an iod, do not initiate 
+	 * queue cleaning if curproc is the pageout daemon. if the NFS mount
+	 * is via local loopback, we may put curproc (pagedaemon) to sleep
+	 * waiting for the writes to complete. But the server (ourself)
+	 * may block the write, waiting for its (ie., our) pagedaemon
+	 * to produce clean pages to handle the write: deadlock.
+	 * XXX: start non-loopback mounts straight away?  If "lots free",
+	 * let pagedaemon start loopback writes anyway?
 	 */
-
-	if (gotiod) {
+	if (gotiod && (curproc != uvm.pagedaemon_proc)) {
 
 		/*
 		 * Ensure that the queue never grows too large.