On Thursday 27 January 2011, Eric Dumazet wrote:

Le jeudi 27 janvier 2011 à 13:38 +0100, Arnd Bergmann a écrit :

quoted

diff --git a/net/x25/x25_out.c b/net/x25/x25_out.c
index d00649f..f1a6ff1 100644
--- a/net/x25/x25_out.c
+++ b/net/x25/x25_out.c

@@ -68,8 +68,11 @@ int x25_output(struct sock *sk, struct sk_buff *skb)
              frontlen = skb_headroom(skb);
 
              while (skb->len > 0) {
-                     if ((skbn = sock_alloc_send_skb(sk, frontlen + max_len,
-                                                     noblock, &err)) == NULL){
+                     release_sock(sk);
+                     skbn = sock_alloc_send_skb(sk, frontlen + max_len,
+                                                1, &err);
+                     lock_sock(sk);
+                     if (!skbn) {
                              if (err == -EWOULDBLOCK && noblock){
                                      kfree_skb(skb);
                                      return sent;

This part looks strange :

noblock variable became "const 1 : NOBLOCK"

Why releasing socket if you dont block in sock_alloc_send_skb() ?

Leftover from an earlier version of the patch, thanks for catching this!

Originally, I wrote this as

	long timeo = sock_sndtimeo(sk, noblock)
	do {
		skbn = sock_alloc_send_skb(sk, frontlen + max_len, 1, &err);
		if (skbn)
			break;

		release_sock(sk);
		timeo = sock_wait_for_wmem(sk, timeo);
		lock_sock(sk);
	} while (timeo);

Then I forgot to flip it back after I noticed that other protocols also just
call release_sock/lock_sock around sock_alloc_send_skb.

I think I'd better go over the whole series and see if there are more things
that got slightly broken...

	Arnd