sis900: fix TX completion

[ Upstream commit 8ac8a010 ]

Since commit 605ad7f1 "tcp: refine TSO autosizing",
outbound throughput is dramatically reduced for some connections, as sis900
is doing TX completion within idle states only.

Make TX completion happen after every transmitted packet.

Test:
netperf

before patch:
> netperf -H remote -l -2000000 -- -s 1000000
MIGRATED TCP STREAM TEST from 0.0.0.0 () port 0 AF_INET to 95.223.112.76 () port 0 AF_INET : demo
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380 327680 327680    253.44      0.06

after patch:
> netperf -H remote -l -10000000 -- -s 1000000
MIGRATED TCP STREAM TEST from 0.0.0.0 () port 0 AF_INET to 95.223.112.76 () port 0 AF_INET : demo
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380 327680 327680    5.38       14.89

Thx to Dave Miller and Eric Dumazet for helpful hints
Signed-off-by: Sergej Benilov <sergej.benilov@googlemail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/net/ethernet/sis/sis900.c

@@ -1057,7 +1057,7 @@ sis900_open(struct net_device *net_dev)
 	sis900_set_mode(sis_priv, HW_SPEED_10_MBPS, FDX_CAPABLE_HALF_SELECTED);
 
 	/* Enable all known interrupts by setting the interrupt mask. */
-	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE);
+	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE | TxDESC);
 	sw32(cr, RxENA | sr32(cr));
 	sw32(ier, IE);
 
@@ -1578,7 +1578,7 @@ static void sis900_tx_timeout(struct net_device *net_dev)
 	sw32(txdp, sis_priv->tx_ring_dma);
 
 	/* Enable all known interrupts by setting the interrupt mask. */
-	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE);
+	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE | TxDESC);
 }
 
 /**
@@ -1618,7 +1618,7 @@ sis900_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
 			spin_unlock_irqrestore(&sis_priv->lock, flags);
 			return NETDEV_TX_OK;
 	}
-	sis_priv->tx_ring[entry].cmdsts = (OWN | skb->len);
+	sis_priv->tx_ring[entry].cmdsts = (OWN | INTR | skb->len);
 	sw32(cr, TxENA | sr32(cr));
 
 	sis_priv->cur_tx ++;
@@ -1674,7 +1674,7 @@ static irqreturn_t sis900_interrupt(int irq, void *dev_instance)
 	do {
 		status = sr32(isr);
 
-		if ((status & (HIBERR|TxURN|TxERR|TxIDLE|RxORN|RxERR|RxOK)) == 0)
+		if ((status & (HIBERR|TxURN|TxERR|TxIDLE|TxDESC|RxORN|RxERR|RxOK)) == 0)
 			/* nothing intresting happened */
 			break;
 		handled = 1;
@@ -1684,7 +1684,7 @@ static irqreturn_t sis900_interrupt(int irq, void *dev_instance)
 			/* Rx interrupt */
 			sis900_rx(net_dev);
 
-		if (status & (TxURN | TxERR | TxIDLE))
+		if (status & (TxURN | TxERR | TxIDLE | TxDESC))
 			/* Tx interrupt */
 			sis900_finish_xmit(net_dev);
 
@@ -1896,8 +1896,8 @@ static void sis900_finish_xmit (struct net_device *net_dev)
 
 		if (tx_status & OWN) {
 			/* The packet is not transmitted yet (owned by hardware) !
-			 * Note: the interrupt is generated only when Tx Machine
-			 * is idle, so this is an almost impossible case */
+			 * Note: this is an almost impossible condition
+			 * in case of TxDESC ('descriptor interrupt') */
 			break;
 		}
 
@@ -2473,7 +2473,7 @@ static int sis900_resume(struct pci_dev *pci_dev)
 	sis900_set_mode(sis_priv, HW_SPEED_10_MBPS, FDX_CAPABLE_HALF_SELECTED);
 
 	/* Enable all known interrupts by setting the interrupt mask. */
-	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE);
+	sw32(imr, RxSOVR | RxORN | RxERR | RxOK | TxURN | TxERR | TxIDLE | TxDESC);
 	sw32(cr, RxENA | sr32(cr));
 	sw32(ier, IE);