From: Andy Gospodarek <gospo@redhat.com>
Date: Thu, 25 Feb 2010 01:50:31 -0500
Subject: [net] ixgbe: stop unmapping DMA buffers too early
Message-id: <20100225015031.GX24578@gospo.rdu.redhat.com>
Patchwork-id: 23424
O-Subject: [PATCH RHEL5.5] ixgbe: stop unmapping DMA buffers too early
Bugzilla: 568153
RH-Acked-by: Chris Wright <chrisw@redhat.com>
RH-Acked-by: Stefan Assmann <sassmann@redhat.com>
RH-Acked-by: John Linville <linville@redhat.com>

This is a not-yet-upstream patch that should resolve issues related to
early calls to pci_unmap_single when using RSC (receive-side coalescing)
on 82599 devices.  For a device that supports RSC, the driver will take
multiple rx buffers and copy them into a single, chained skb and pass
that skb up the stack.  Because the hardware may still have some writing
to do in the DMA area associated with the first skb in the chain, we
need to wait to unmap it until all buffers associated with that skb are
collected and unmapped.

This patch came from Intel has plans to post it as it appears here 'in
the next few days' as of 24 Feb.

This has been verified by Intel, Chris Wright, Mark Wagner and me.

This will resolve RHBZ 568153.


diff --git a/drivers/net/ixgbe/ixgbe_main.c b/drivers/net/ixgbe/ixgbe_main.c
index 1f76a12..7e52edb 100644
--- a/drivers/net/ixgbe/ixgbe_main.c
+++ b/drivers/net/ixgbe/ixgbe_main.c
@@ -821,6 +821,12 @@ static inline struct sk_buff *ixgbe_transform_rsc_queue(struct sk_buff *skb,
 	return skb;
 }
 
+struct ixgbe_rsc_cb {
+	dma_addr_t dma;
+};
+
+#define IXGBE_RSC_CB(skb) ((struct ixgbe_rsc_cb *)(skb)->cb)
+
 static bool ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
                                struct ixgbe_ring *rx_ring,
                                int *work_done, int work_to_do)
@@ -869,9 +875,21 @@ static bool ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 		rx_buffer_info->skb = NULL;
 
 		if (rx_buffer_info->dma) {
-			pci_unmap_single(pdev, rx_buffer_info->dma,
-			                 rx_ring->rx_buf_len,
-			                 PCI_DMA_FROMDEVICE);
+			if ((adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED) &&
+			    (!(staterr & IXGBE_RXD_STAT_EOP)) &&
+				 (!(skb->prev)))
+				/*
+				 * When HWRSC is enabled, delay unmapping
+				 * of the first packet. It carries the
+				 * header information, HW may still
+				 * access the header after the writeback.
+				 * Only unmap it when EOP is reached
+				 */
+				IXGBE_RSC_CB(skb)->dma = rx_buffer_info->dma;
+			else
+				pci_unmap_single(pdev, rx_buffer_info->dma,
+				                 rx_ring->rx_buf_len,
+				                 PCI_DMA_FROMDEVICE);
 			rx_buffer_info->dma = 0;
 			skb_put(skb, len);
 		}
@@ -917,10 +935,17 @@ static bool ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 
 		if (staterr & IXGBE_RXD_STAT_EOP) {
 			if (skb->prev)
-				skb = ixgbe_transform_rsc_queue(skb, &(rx_ring->rsc_count));
+				skb = ixgbe_transform_rsc_queue(skb,
+						&(rx_ring->rsc_count));
 			if (adapter->flags2 & IXGBE_FLAG2_RSC_ENABLED) {
+				if (IXGBE_RSC_CB(skb)->dma)
+					pci_unmap_single(pdev,
+							 IXGBE_RSC_CB(skb)->dma,
+							 rx_ring->rx_buf_len,
+							 PCI_DMA_FROMDEVICE);
 				if (rx_ring->flags & IXGBE_RING_RX_PS_ENABLED)
-					rx_ring->rsc_count += skb_shinfo(skb)->nr_frags;
+					rx_ring->rsc_count +=
+						skb_shinfo(skb)->nr_frags;
 				else
 					rx_ring->rsc_count++;
 				rx_ring->rsc_flush++;
@@ -3133,6 +3158,10 @@ static void ixgbe_clean_rx_ring(struct ixgbe_adapter *adapter,
 			rx_buffer_info->skb = NULL;
 			do {
 				struct sk_buff *this = skb;
+				if (IXGBE_RSC_CB(this)->dma)
+					pci_unmap_single(pdev, IXGBE_RSC_CB(this)->dma,
+					                 rx_ring->rx_buf_len,
+					                 PCI_DMA_FROMDEVICE);
 				skb = skb->prev;
 				dev_kfree_skb(this);
 			} while (skb);