Back out smc91x late collision "performance" hack.

Apparantly, the reason for this change is that it gives "much less
NFS stalls with a SMC91C96 on a 10/100MB switch."

Alan Cox's comments on this is:
"spec says 16 collision retries max _ever_.  I bet his 'broken
switch' is a case of half v full duplex" and a recommendation to
remove it.

Also, since the change does not limit the number of times we may
retry to send the packet, I think this code opens a potential DoS
attack against network which has a machine connected running this
code.

drivers/net/smc91x.c

@@ -693,17 +693,19 @@ static void smc_tx(struct net_device *dev)
 	if (tx_status & TS_LOSTCAR)
 		lp->stats.tx_carrier_errors++;
 
-	SMC_WAIT_MMU_BUSY();
-
 	if (tx_status & TS_LATCOL) {
 		PRINTK("%s: late collision occurred on last xmit\n", dev->name);
 		lp->stats.tx_window_errors++;
-		/* It's really cheap to requeue the pkt here */
-		SMC_SET_MMU_CMD( MC_ENQUEUE );
-	} else {
+		if (!(lp->stats.tx_window_errors & 63) && net_ratelimit()) {
+			printk(KERN_INFO "%s: unexpectedly large numbers of "
+			       "late collisions. Please check duplex "
+			       "setting.\n", dev->name);
+		}
+	}
+
 	/* kill the packet */
+	SMC_WAIT_MMU_BUSY();
 	SMC_SET_MMU_CMD(MC_FREEPKT);
-	}
 
 	/* Don't restore Packet Number Reg until busy bit is cleared */
 	SMC_WAIT_MMU_BUSY();