r8169: improve rtl8169_start_xmit

Improve the following in rtl8169_start_xmit:
- tp->cur_tx can be accessed in parallel by rtl_tx(), therefore
  annotate the race by using WRITE_ONCE
- avoid checking stop_queue a second time by moving the doorbell check
- netif_stop_queue() uses atomic operation set_bit() that includes a
  full memory barrier on some platforms, therefore use
  smp_mb__after_atomic to avoid overhead

Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Link: https://lore.kernel.org/r/80085451-3eaf-507a-c7c0-08d607c46fbc@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/realtek/r8169_main.c

@@ -4226,7 +4226,7 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 	/* rtl_tx needs to see descriptor changes before updated tp->cur_tx */
 	smp_wmb();
 
-	tp->cur_tx += frags + 1;
+	WRITE_ONCE(tp->cur_tx, tp->cur_tx + frags + 1);
 
 	stop_queue = !rtl_tx_slots_avail(tp, MAX_SKB_FRAGS);
 	if (unlikely(stop_queue)) {
@@ -4235,25 +4235,22 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 		 */
 		smp_wmb();
 		netif_stop_queue(dev);
-		door_bell = true;
-	}
-
-	if (door_bell)
-		rtl8169_doorbell(tp);
-
-	if (unlikely(stop_queue)) {
 		/* Sync with rtl_tx:
 		 * - publish queue status and cur_tx ring index (write barrier)
 		 * - refresh dirty_tx ring index (read barrier).
 		 * May the current thread have a pessimistic view of the ring
 		 * status and forget to wake up queue, a racing rtl_tx thread
 		 * can't.
 		 */
-		smp_mb();
+		smp_mb__after_atomic();
 		if (rtl_tx_slots_avail(tp, MAX_SKB_FRAGS))
 			netif_start_queue(dev);
+		door_bell = true;
 	}
 
+	if (door_bell)
+		rtl8169_doorbell(tp);
+
 	return NETDEV_TX_OK;
 
 err_dma_1: