From: Peter Staubach <staubach@redhat.com>
Date: Mon, 9 Jun 2008 15:14:40 -0400
Subject: [nfs] ensure that options turn off attribute caching
Message-id: 484D8120.4070300@redhat.com
O-Subject: [RHEL-5.3 PATCH] Ensure that 'noac' and/or 'actimeo=0' turn off attribute caching
Bugzilla: 450184
RH-Acked-by: Steve Dickson <SteveD@redhat.com>
RH-Acked-by: Jeff Layton <jlayton@redhat.com>

Hi.

Attached is a patch to address bz450184, "Ensure that 'noac' and/or
'actimeo=0' turn off attribute caching".

The problem is that the mount options, "noac" and/or "actimeo=0",
are supposed to stop the NFS client from doing attribute caching.
This means that when the attributes of a file are requested, they
are retrieved from the server instead of being satisfied locally
on the client.

However, the attribute cache timeout is based on jiffies.  The
cache timeout algorithm works by comparing the current jiffies
value against the sum of the jiffies value that was stored when
the attributes were previously retrieved and the cache duration
time value.  If the current jiffies value is greater than that
sum, then the attribute cache is deemed to be timed out and a
new set of attributes are retrieved from the server.

Unfortunately, this allows the attributes to be cached for 1
jiffie longer than the cache timeout is supposed to be.  For the
most part, this does not matter.  Cache timeouts are normally
expressed in seconds and the jiffies values are usually in units
1/1000 of a second, so the additional jiffie isn't a big deal.
However, when the timeout is supposed to be 0, then the additional
jiffie can be significant.

Some applications, notably Oracle 10G, depend upon being able to
detect file attribute changes, all of which can occur during the
same jiffies value.  They use one or the other of the two listed
NFS mount options in order to be able to detect those changes.
However, since the attribute cache timeouts are off by one, then
these applications can not correctly detect the file attribute
changes in a timely enough fashion.

This patch was constructed by Trond, the upstream maintainer.

    Comments?

       ps

diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index 53f96fa..ff11522 100644
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -671,6 +671,13 @@ int nfs_attribute_timeout(struct inode *inode)
 
 	if (nfs_have_delegation(inode, FMODE_READ))
 		return 0;
+	/*
+	 * Special case: if the attribute timeout is set to 0, then we
+	 *               treat the cache as having expired (unless we
+	 *               have a delegation).
+	 */
+	if (nfsi->attrtimeo == 0)
+		return 1;
 	return time_after(jiffies, nfsi->read_cache_jiffies+nfsi->attrtimeo);
 }