[PATCH] 2.4.4 linearize UDP RPC requests using GFP_KERNEL...

[Trond Myklebust]

Hi,

  The zero-copy TCP patch that was integrated in 2.4.4 has allowed us
to move the task of reassembling of the fragments into one buffer
until we're out of the ->data_ready() bottom half context.
  In 2.4.4 therefore, the nfsd thread can end up calling the function
skb_linearize(), in order to allocate a new buffer and copy over the
fragments.

  IMHO allocating the buffer using GFP_ATOMIC is a mistake. As I said
we're in a thread context, so sleeping in GFP_KERNEL is safe. In
addition, the cost of dropping the request if we can't allocate the
buffer is heavy in that the client has to wait for a timeout, and then
retry.

I'd therefore like to propose the following change.

Cheers,
  Trond

--- linux-2.4.4/net/sunrpc/svcsock.c.orig	Fri Apr 27 23:15:01 2001
+++ linux-2.4.4/net/sunrpc/svcsock.c	Fri May 11 10:08:36 2001
@@ -383,7 +383,7 @@
 
 	/* Sorry. */
 	if (skb_is_nonlinear(skb)) {
-		if (skb_linearize(skb, GFP_ATOMIC) != 0) {
+		if (skb_linearize(skb, GFP_KERNEL) != 0) {
 			kfree_skb(skb);
 			svc_sock_received(svsk, 0);
 			return 0;