NFS clients hang

Hardware:
Dell Precision 530 Dual Xeon1.7GHz 1GB RDRAM

Redhat 7.2 distribution is used with kernel 2.4.21 patched with IngoMolnar IRQ balancing as there is a 'bug' with Xeon processors (only one would take care of all interrupts), configured with CONFIG_HIGHMEM option. nfs-utils-1.0.3-1 is used. The mount option for the client is: rw,nosuid,nodev,hard,intr,bg,rsize=8192,wsize=8192,nfsvers=3,lock,udp

Sometimes the NFS client was hanging for no reason (usually when a medium size file (30MB+) was handled over NFS or a lot files were moved around).
The machine was then completely locked (no access at all to it, even thru serial console, no display on the monitor anymore). tcpdump files didn't show anything wrong but just stop to report traffic until the lock. I reconfigured the kernel with CONFIG_DEBUG_SPINLOCK and CONFIG_SOFT_WATCHDOG to have access to some debug info. The client was still completely locked up and couldn't get out of the locked state.

I finally read in Chuck Lever's technical report "Customers that use 2.4 kernels on hardware with more than 896MB shoould know that a special kernel option, known as CONFIG_HIGHMEM, is required to access and use memory above 896M. The linux NFS client has a known problem in these configurations where an application or the whole client system can hang at random. This issue has been addressed in the 2.4.20 kernel, but still haunts kernels contained in distribution from RedHat and SUSE that are based on earlier kernels".

So I recompiled 2.4.21 without CONFIG_HIGHMEM and since the NFS clients have been rock solid. Not a single hang since but of course 100+MB of memory are 'lost' now. So is this problem still around or is it specific to Xeons?
Thanks

JC

-------------------------------------------------------
This SF.Net email sponsored by: Parasoft
Error proof Web apps, automate testing & more.
Download & eval WebKing and get a free book.
www.parasoft.com/bulletproofapps1
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

> So is this problem still around or is it specific to Xeons?

It was believed to be fixed in the final version of 2.4.21. Andrea Archangeli made a patch that adds a non-blocking version of the kmap() call.

Hmm... It all looks very wrong though: We certainly don't want to be sending partial requests...
Does the following patch help?

Cheers,
Trond

diff -u --recursive --new-file linux-2.4.22-pre4/net/sunrpc/xdr.c linux-2.4.22-fix_highmem/net/sunrpc/xdr.c
@@ -180,7 +180,8 @@
 {
 	struct iovec	*iov = iov_base;
 	struct page	**ppage = xdr->pages;
-	unsigned int	len, pglen = xdr->page_len, first_kmap;
+	struct page	**first_kmap = NULL;
+	unsigned int	len, pglen = xdr->page_len;
 
 	len = xdr->head[0].iov_len;
 	if (base < len) {

@@ -203,16 +204,15 @@
 		ppage += base >> PAGE_CACHE_SHIFT;
 		base &= ~PAGE_CACHE_MASK;
 	}
-	first_kmap = 1;
 	do {
 		len = PAGE_CACHE_SIZE;
-		if (first_kmap) {
-			first_kmap = 0;
+		if (!first_kmap) {
+			first_kmap = ppage;
 			iov->iov_base = kmap(*ppage);
 		} else {
 			iov->iov_base = kmap_nonblock(*ppage);
 			if (!iov->iov_base)
-				goto out;
+				goto out_err;
 		}
 		if (base) {
 			iov->iov_base += base;

@@ -233,6 +233,10 @@
 	}
  out:
 	return (iov - iov_base);
+out_err:
+	for (; first_kmap != ppage; first_kmap++)
+		kunmap(*first_kmap);
+	return 0;
 }
 
 void xdr_kunmap(struct xdr_buf *xdr, unsigned int base, int niov)

diff -u --recursive --new-file linux-2.4.22-pre4/net/sunrpc/xprt.c linux-2.4.22-fix_highmem/net/sunrpc/xprt.c
@@ -237,6 +237,10 @@
 		unsigned int slen_part, n;
 
 		niov = xdr_kmap(niv, xdr, skip);
+		if (!niov) {
+			result = -EAGAIN;
+			break;
+		}
 
 		msg.msg_flags   = MSG_DONTWAIT|MSG_NOSIGNAL;
 		msg.msg_iov	= niv;

-------------------------------------------------------
This SF.Net email sponsored by: Parasoft
Error proof Web apps, automate testing & more.
Download & eval WebKing and get a free book.
www.parasoft.com/bulletproofapps1
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs