[net-next-2.6 PATCH 1/4] e1000e: save skb counts in TX to avoid cache misses

[net-next-2.6 PATCH 1/4] e1000e: save skb counts in TX to avoid cache misses

[Jeff Kirsher]

From: Tom Herbert <therbert@google.com>

In e1000_tx_map, precompute number of segements and bytecounts which
are derived from fields in skb; these are stored in buffer_info.  When
cleaning tx in e1000_clean_tx_irq use the values in the associated
buffer_info for statistics counting, this eliminates cache misses
on skb fields.

Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
---

 drivers/net/e1000e/e1000.h  |    2 ++
 drivers/net/e1000e/netdev.c |   18 +++++++++---------
 2 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/drivers/net/e1000e/e1000.h b/drivers/net/e1000e/e1000.h
index 906c4da..c0b3db4 100644
--- a/drivers/net/e1000e/e1000.h
+++ b/drivers/net/e1000e/e1000.h
@@ -189,6 +189,8 @@ struct e1000_buffer {
 			unsigned long time_stamp;
 			u16 length;
 			u16 next_to_watch;
+			unsigned int segs;
+			unsigned int bytecount;
 			u16 mapped_as_page;
 		};
 		/* Rx */
diff --git a/drivers/net/e1000e/netdev.c b/drivers/net/e1000e/netdev.c
index 0a16465..1eb9b59 100644
--- a/drivers/net/e1000e/netdev.c
+++ b/drivers/net/e1000e/netdev.c
@@ -1001,14 +1001,8 @@ static bool e1000_clean_tx_irq(struct e1000_adapter *adapter)
 			cleaned = (i == eop);
 
 			if (cleaned) {
-				struct sk_buff *skb = buffer_info->skb;
-				unsigned int segs, bytecount;
-				segs = skb_shinfo(skb)->gso_segs ?: 1;
-				/* multiply data chunks by size of headers */
-				bytecount = ((segs - 1) * skb_headlen(skb)) +
-					    skb->len;
-				total_tx_packets += segs;
-				total_tx_bytes += bytecount;
+				total_tx_packets += buffer_info->segs;
+				total_tx_bytes += buffer_info->bytecount;
 			}
 
 			e1000_put_txbuf(adapter, buffer_info);
@@ -4277,7 +4271,7 @@ static int e1000_tx_map(struct e1000_adapter *adapter,
 	struct e1000_buffer *buffer_info;
 	unsigned int len = skb_headlen(skb);
 	unsigned int offset = 0, size, count = 0, i;
-	unsigned int f;
+	unsigned int f, bytecount, segs;
 
 	i = tx_ring->next_to_use;
 
@@ -4337,7 +4331,13 @@ static int e1000_tx_map(struct e1000_adapter *adapter,
 		}
 	}
 
+	segs = skb_shinfo(skb)->gso_segs ?: 1;
+	/* multiply data chunks by size of headers */
+	bytecount = ((segs - 1) * skb_headlen(skb)) + skb->len;
+
 	tx_ring->buffer_info[i].skb = skb;
+	tx_ring->buffer_info[i].segs = segs;
+	tx_ring->buffer_info[i].bytecount = bytecount;
 	tx_ring->buffer_info[first].next_to_watch = i;
 
 	return count;

[net-next-2.6 PATCH 2/4] e1000e: reduce writes of RX producer ptr

[Jeff Kirsher]

From: Tom Herbert <therbert@google.com>

Reduce number of writes to RX producer pointer.   When alloc'ing RX
buffers, only write the RX producer pointer once every
E1000_RX_BUFFER_WRITE (16) buffers created.

Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
---

 drivers/net/e1000e/netdev.c |   57 +++++++++++++++++--------------------------
 1 files changed, 23 insertions(+), 34 deletions(-)

diff --git a/drivers/net/e1000e/netdev.c b/drivers/net/e1000e/netdev.c
index 1eb9b59..b767dbb 100644
--- a/drivers/net/e1000e/netdev.c
+++ b/drivers/net/e1000e/netdev.c
@@ -548,26 +548,23 @@ map_skb:
 		rx_desc = E1000_RX_DESC(*rx_ring, i);
 		rx_desc->buffer_addr = cpu_to_le64(buffer_info->dma);
 
+		if (unlikely(!(i & (E1000_RX_BUFFER_WRITE - 1)))) {
+			/*
+			 * Force memory writes to complete before letting h/w
+			 * know there are new descriptors to fetch.  (Only
+			 * applicable for weak-ordered memory model archs,
+			 * such as IA-64).
+			 */
+			wmb();
+			writel(i, adapter->hw.hw_addr + rx_ring->tail);
+		}
 		i++;
 		if (i == rx_ring->count)
 			i = 0;
 		buffer_info = &rx_ring->buffer_info[i];
 	}
 
-	if (rx_ring->next_to_use != i) {
-		rx_ring->next_to_use = i;
-		if (i-- == 0)
-			i = (rx_ring->count - 1);
-
-		/*
-		 * Force memory writes to complete before letting h/w
-		 * know there are new descriptors to fetch.  (Only
-		 * applicable for weak-ordered memory model archs,
-		 * such as IA-64).
-		 */
-		wmb();
-		writel(i, adapter->hw.hw_addr + rx_ring->tail);
-	}
+	rx_ring->next_to_use = i;
 }
 
 /**
@@ -649,6 +646,17 @@ static void e1000_alloc_rx_buffers_ps(struct e1000_adapter *adapter,
 
 		rx_desc->read.buffer_addr[0] = cpu_to_le64(buffer_info->dma);
 
+		if (unlikely(!(i & (E1000_RX_BUFFER_WRITE - 1)))) {
+			/*
+			 * Force memory writes to complete before letting h/w
+			 * know there are new descriptors to fetch.  (Only
+			 * applicable for weak-ordered memory model archs,
+			 * such as IA-64).
+			 */
+			wmb();
+			writel(i<<1, adapter->hw.hw_addr + rx_ring->tail);
+		}
+
 		i++;
 		if (i == rx_ring->count)
 			i = 0;
@@ -656,26 +664,7 @@ static void e1000_alloc_rx_buffers_ps(struct e1000_adapter *adapter,
 	}
 
 no_buffers:
-	if (rx_ring->next_to_use != i) {
-		rx_ring->next_to_use = i;
-
-		if (!(i--))
-			i = (rx_ring->count - 1);
-
-		/*
-		 * Force memory writes to complete before letting h/w
-		 * know there are new descriptors to fetch.  (Only
-		 * applicable for weak-ordered memory model archs,
-		 * such as IA-64).
-		 */
-		wmb();
-		/*
-		 * Hardware increments by 16 bytes, but packet split
-		 * descriptors are 32 bytes...so we increment tail
-		 * twice as much.
-		 */
-		writel(i<<1, adapter->hw.hw_addr + rx_ring->tail);
-	}
+	rx_ring->next_to_use = i;
 }
 
 /**

Re: [net-next-2.6 PATCH 2/4] e1000e: reduce writes of RX producer ptr

[David Miller]

From: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Date: Wed, 05 May 2010 17:02:49 -0700

> From: Tom Herbert <therbert@google.com>
> 
> Reduce number of writes to RX producer pointer.   When alloc'ing RX
> buffers, only write the RX producer pointer once every
> E1000_RX_BUFFER_WRITE (16) buffers created.
> 
> Signed-off-by: Tom Herbert <therbert@google.com>
> Acked-by: Bruce Allan <bruce.w.allan@intel.com>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Applied.

[net-next-2.6 PATCH 3/4] e1000e: Remove unnessary log message

[Jeff Kirsher]

From: Tom Herbert <therbert@google.com>

Remove e_info message printed whenever TSO is enabled or disabled.
This is not very useful and just clutters dmesg.

Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
---

 drivers/net/e1000e/ethtool.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

diff --git a/drivers/net/e1000e/ethtool.c b/drivers/net/e1000e/ethtool.c
index c81118a..6ff376c 100644
--- a/drivers/net/e1000e/ethtool.c
+++ b/drivers/net/e1000e/ethtool.c
@@ -412,7 +412,6 @@ static int e1000_set_tso(struct net_device *netdev, u32 data)
 		netdev->features &= ~NETIF_F_TSO6;
 	}
 
-	e_info("TSO is %s\n", data ? "Enabled" : "Disabled");
 	adapter->flags |= FLAG_TSO_FORCE;
 	return 0;
 }

Re: [net-next-2.6 PATCH 3/4] e1000e: Remove unnessary log message

[David Miller]

From: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Date: Wed, 05 May 2010 17:03:11 -0700

> From: Tom Herbert <therbert@google.com>
> 
> Remove e_info message printed whenever TSO is enabled or disabled.
> This is not very useful and just clutters dmesg.
> 
> Signed-off-by: Tom Herbert <therbert@google.com>
> Acked-by: Bruce Allan <bruce.w.allan@intel.com>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Applied.

[net-next-2.6 PATCH 4/4] e1000e: Save irq into netdev structure

[Jeff Kirsher]

From: Tom Herbert <therbert@google.com>

Set net->devirq to pdev->irq.  This should be consistent with other
drivers.

Signed-off-by: Tom Herbert <therbert@google.com>
Acked-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
---

 drivers/net/e1000e/netdev.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/drivers/net/e1000e/netdev.c b/drivers/net/e1000e/netdev.c
index b767dbb..c5f65a2 100644
--- a/drivers/net/e1000e/netdev.c
+++ b/drivers/net/e1000e/netdev.c
@@ -5456,6 +5456,8 @@ static int __devinit e1000_probe(struct pci_dev *pdev,
 
 	SET_NETDEV_DEV(netdev, &pdev->dev);
 
+	netdev->irq = pdev->irq;
+
 	pci_set_drvdata(pdev, netdev);
 	adapter = netdev_priv(netdev);
 	hw = &adapter->hw;

Re: [net-next-2.6 PATCH 4/4] e1000e: Save irq into netdev structure

[David Miller]

From: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Date: Wed, 05 May 2010 17:03:32 -0700

> From: Tom Herbert <therbert@google.com>
> 
> Set net->devirq to pdev->irq.  This should be consistent with other
> drivers.
> 
> Signed-off-by: Tom Herbert <therbert@google.com>
> Acked-by: Bruce Allan <bruce.w.allan@intel.com>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Applied.

Re: [net-next-2.6 PATCH 1/4] e1000e: save skb counts in TX to avoid cache misses

[David Miller]

From: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Date: Wed, 05 May 2010 17:02:27 -0700

> From: Tom Herbert <therbert@google.com>
> 
> In e1000_tx_map, precompute number of segements and bytecounts which
> are derived from fields in skb; these are stored in buffer_info.  When
> cleaning tx in e1000_clean_tx_irq use the values in the associated
> buffer_info for statistics counting, this eliminates cache misses
> on skb fields.
> 
> Signed-off-by: Tom Herbert <therbert@google.com>
> Acked-by: Bruce Allan <bruce.w.allan@intel.com>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Applied.