Fix a subtle race condition between tg3_start_xmit() and tg3_tx()
discovered by Herbert Xu <herbert@gondor.apana.org.au>:

CPU0					CPU1
tg3_start_xmit()
	if (tx_ring_full) {
		tx_lock
					tg3_tx()
						if (!netif_queue_stopped)
		netif_stop_queue()
		if (!tx_ring_full)
						update_tx_ring 
			netif_wake_queue()
		tx_unlock
	}

Even though tx_ring is updated before the if statement in tg3_tx() in
program order, it can be re-ordered by the CPU as shown above.  This
scenario can cause the tx queue to be stopped forever if tg3_tx() has
just freed up the entire tx_ring.  The possibility of this happening
should be very rare though.

The following changes are made:

1. Add memory barrier to fix the above race condition.

2. Eliminate the private tx_lock altogether and rely solely on
netif_tx_lock.  This eliminates one spinlock in tg3_start_xmit()
when the ring is full.

3. Because of 2, use netif_tx_lock in tg3_tx() before calling
netif_wake_queue().

4. Change TX_BUFFS_AVAIL to an inline function with a memory barrier.
Herbert and David suggested using the memory barrier instead of
volatile.

5. Check for the full wake queue condition before getting
netif_tx_lock in tg3_tx().  This reduces the number of unnecessary
spinlocks when the tx ring is full in a steady-state condition.

6. Update version to 3.65.
Signed-off-by: Michael Chan <mchan@broadcom.com>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>