tun: do not put self in waitq if doing a nonblock read

Perf shows a relatively high rate (about 8%) race in
spin_lock_irqsave() when doing netperf between external host and
guest. It's mainly becuase the lock contention between the
tun_do_read() and tun_xmit_skb(), so this patch do not put self into
waitqueue to reduce this kind of race. After this patch, it drops to
4%.

Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Amos Kong <akong@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/tun.c

@@ -817,7 +817,8 @@ static ssize_t tun_do_read(struct tun_struct *tun,
 
 	tun_debug(KERN_INFO, tun, "tun_chr_read\n");
 
-	add_wait_queue(&tun->wq.wait, &wait);
+	if (unlikely(!noblock))
+		add_wait_queue(&tun->wq.wait, &wait);
 	while (len) {
 		current->state = TASK_INTERRUPTIBLE;
 
@@ -848,7 +849,8 @@ static ssize_t tun_do_read(struct tun_struct *tun,
 	}
 
 	current->state = TASK_RUNNING;
-	remove_wait_queue(&tun->wq.wait, &wait);
+	if (unlikely(!noblock))
+		remove_wait_queue(&tun->wq.wait, &wait);
 
 	return ret;
 }