tg3: Improve small packet performance

smp_mb() inside tg3_tx_avail() is used twice in the normal
tg3_start_xmit() path (see illustration below).  The full memory
barrier is only necessary during race conditions with tx completion.
We can speed up the tx path by replacing smp_mb() in tg3_tx_avail()
with a compiler barrier.  The compiler barrier is to force the
compiler to fetch the tx_prod and tx_cons from memory.

In the race condition between tg3_start_xmit() and tg3_tx(),
we have the following situation:

tg3_start_xmit()                       tg3_tx()
    if (!tg3_tx_avail())
        BUG();

    ...

    if (!tg3_tx_avail())
        netif_tx_stop_queue();         update_tx_index();
        smp_mb();                      smp_mb();
        if (tg3_tx_avail())            if (netif_tx_queue_stopped() &&
            netif_tx_wake_queue();         tg3_tx_avail())

With smp_mb() removed from tg3_tx_avail(), we need to add smp_mb() to
tg3_start_xmit() as shown above to properly order netif_tx_stop_queue()
and tg3_tx_avail() to check the ring index.  If it is not strictly
ordered, the tx queue can be stopped forever.

This improves performance by about 3% with 2 ports running
bi-directional 64-byte packets.
Reviewed-by: Benjamin Li <benli@broadcom.com>
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: Matt Carlson <mcarlson@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/tg3.c

@@ -4386,7 +4386,8 @@ static void tg3_tx_recover(struct tg3 *tp)
 
 static inline u32 tg3_tx_avail(struct tg3_napi *tnapi)
 {
-	smp_mb();
+	/* Tell compiler to fetch tx indices from memory. */
+	barrier();
 	return tnapi->tx_pending -
 	       ((tnapi->tx_prod - tnapi->tx_cons) & (TG3_TX_RING_SIZE - 1));
 }
@@ -5670,6 +5671,13 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb,
 	tnapi->tx_prod = entry;
 	if (unlikely(tg3_tx_avail(tnapi) <= (MAX_SKB_FRAGS + 1))) {
 		netif_tx_stop_queue(txq);
+
+		/* netif_tx_stop_queue() must be done before checking
+		 * checking tx index in tg3_tx_avail() below, because in
+		 * tg3_tx(), we update tx index before checking for
+		 * netif_tx_queue_stopped().
+		 */
+		smp_mb();
 		if (tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi))
 			netif_tx_wake_queue(txq);
 	}
@@ -5715,6 +5723,13 @@ static int tg3_tso_bug(struct tg3 *tp, struct sk_buff *skb)
 	/* Estimate the number of fragments in the worst case */
 	if (unlikely(tg3_tx_avail(&tp->napi[0]) <= frag_cnt_est)) {
 		netif_stop_queue(tp->dev);
+
+		/* netif_tx_stop_queue() must be done before checking
+		 * checking tx index in tg3_tx_avail() below, because in
+		 * tg3_tx(), we update tx index before checking for
+		 * netif_tx_queue_stopped().
+		 */
+		smp_mb();
 		if (tg3_tx_avail(&tp->napi[0]) <= frag_cnt_est)
 			return NETDEV_TX_BUSY;
 
@@ -5950,6 +5965,13 @@ static netdev_tx_t tg3_start_xmit_dma_bug(struct sk_buff *skb,
 	tnapi->tx_prod = entry;
 	if (unlikely(tg3_tx_avail(tnapi) <= (MAX_SKB_FRAGS + 1))) {
 		netif_tx_stop_queue(txq);
+
+		/* netif_tx_stop_queue() must be done before checking
+		 * checking tx index in tg3_tx_avail() below, because in
+		 * tg3_tx(), we update tx index before checking for
+		 * netif_tx_queue_stopped().
+		 */
+		smp_mb();
 		if (tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi))
 			netif_tx_wake_queue(txq);
 	}