Skip to contentSkip to conversation

mlx4_en: Saving mem access on data path

PATCH In review Yevgeny Petrilin 4 participants 5 messages Feb 23, 2012 – Mar 1, 2012

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 3/3 mlx4_en: Saving mem access on data path

Localized the pdev->dev, and using dma_map instead of pci_map There are multiple map/unmap operations on data path, optimizing those by saving redundant pointer access.

---
 drivers/net/ethernet/mellanox/mlx4/en_netdev.c |    1 +
 drivers/net/ethernet/mellanox/mlx4/en_rx.c     |   14 +++++---------
 drivers/net/ethernet/mellanox/mlx4/en_tx.c     |   13 ++++++-------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h   |    1 +
 4 files changed, 13 insertions(+), 16 deletions(-)
drivers/net/ethernet/mellanox/mlx4/en_netdev.c +1 -0
index 9fe4f94..31b455a 100644
@@ -1062,6 +1062,7 @@ int mlx4_en_init_netdev(struct mlx4_en_dev *mdev, int port,
 	memset(priv, 0, sizeof(struct mlx4_en_priv));
 	priv->dev = dev;
 	priv->mdev = mdev;
+	priv->ddev = &mdev->pdev->dev;
 	priv->prof = prof;
 	priv->port = port;
 	priv->port_up = false;
drivers/net/ethernet/mellanox/mlx4/en_rx.c +5 -9
index d703ef2..c881712 100644
@@ -48,7 +48,6 @@ static int mlx4_en_alloc_frag(struct mlx4_en_priv *priv,
 			      struct mlx4_en_rx_alloc *ring_alloc,
 			      int i)
 {
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
 	struct mlx4_en_rx_alloc *page_alloc = &ring_alloc[i];
 	struct page *page;
@@ -72,7 +71,7 @@ static int mlx4_en_alloc_frag(struct mlx4_en_priv *priv,
 		skb_frags[i].offset = page_alloc->offset;
 		page_alloc->offset += frag_info->frag_stride;
 	}
-	dma = pci_map_single(mdev->pdev, page_address(skb_frags[i].page) +
+	dma = dma_map_single(priv->ddev, page_address(skb_frags[i].page) +
 			     skb_frags[i].offset, frag_info->frag_size,
 			     PCI_DMA_FROMDEVICE);
 	rx_desc->data[i].addr = cpu_to_be64(dma);
@@ -186,7 +185,6 @@ static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
 				 struct mlx4_en_rx_ring *ring,
 				 int index)
 {
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct page_frag *skb_frags;
 	struct mlx4_en_rx_desc *rx_desc = ring->buf + (index << ring->log_stride);
 	dma_addr_t dma;
@@ -198,7 +196,7 @@ static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
 		dma = be64_to_cpu(rx_desc->data[nr].addr);
 
 		en_dbg(DRV, priv, "Unmapping buffer at dma:0x%llx\n", (u64) dma);
-		pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
+		dma_unmap_single(priv->ddev, dma, skb_frags[nr].size,
 				 PCI_DMA_FROMDEVICE);
 		put_page(skb_frags[nr].page);
 	}
@@ -412,7 +410,6 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 				    int length)
 {
 	struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct mlx4_en_frag_info *frag_info;
 	int nr;
 	dma_addr_t dma;
@@ -435,7 +432,7 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 			goto fail;
 
 		/* Unmap buffer */
-		pci_unmap_single(mdev->pdev, dma, skb_frag_size(&skb_frags_rx[nr]),
+		dma_unmap_single(priv->ddev, dma, skb_frag_size(&skb_frags_rx[nr]),
 				 PCI_DMA_FROMDEVICE);
 	}
 	/* Adjust size of last fragment to match actual length */
@@ -461,7 +458,6 @@ static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
 				      struct mlx4_en_rx_alloc *page_alloc,
 				      unsigned int length)
 {
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct sk_buff *skb;
 	void *va;
 	int used_frags;
@@ -483,10 +479,10 @@ static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
 		/* We are copying all relevant data to the skb - temporarily
 		 * synch buffers for the copy */
 		dma = be64_to_cpu(rx_desc->data[0].addr);
-		dma_sync_single_for_cpu(&mdev->pdev->dev, dma, length,
+		dma_sync_single_for_cpu(priv->ddev, dma, length,
 					DMA_FROM_DEVICE);
 		skb_copy_to_linear_data(skb, va, length);
-		dma_sync_single_for_device(&mdev->pdev->dev, dma, length,
+		dma_sync_single_for_device(priv->ddev, dma, length,
 					   DMA_FROM_DEVICE);
 		skb->tail += length;
 	} else {
drivers/net/ethernet/mellanox/mlx4/en_tx.c +6 -7
index e452bba..829d0cf 100644
@@ -198,7 +198,6 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 				struct mlx4_en_tx_ring *ring,
 				int index, u8 owner)
 {
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
@@ -214,7 +213,7 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
 		if (!tx_info->inl) {
 			if (tx_info->linear) {
-				pci_unmap_single(mdev->pdev,
+				dma_unmap_single(priv->ddev,
 					(dma_addr_t) be64_to_cpu(data->addr),
 					 be32_to_cpu(data->byte_count),
 					 PCI_DMA_TODEVICE);
@@ -223,7 +222,7 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 
 			for (i = 0; i < frags; i++) {
 				frag = &skb_shinfo(skb)->frags[i];
-				pci_unmap_page(mdev->pdev,
+				dma_unmap_page(priv->ddev,
 					(dma_addr_t) be64_to_cpu(data[i].addr),
 					skb_frag_size(frag), PCI_DMA_TODEVICE);
 			}
@@ -241,7 +240,7 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 			}
 
 			if (tx_info->linear) {
-				pci_unmap_single(mdev->pdev,
+				dma_unmap_single(priv->ddev,
 					(dma_addr_t) be64_to_cpu(data->addr),
 					 be32_to_cpu(data->byte_count),
 					 PCI_DMA_TODEVICE);
@@ -253,7 +252,7 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 				if ((void *) data >= end)
 					data = ring->buf;
 				frag = &skb_shinfo(skb)->frags[i];
-				pci_unmap_page(mdev->pdev,
+				dma_unmap_page(priv->ddev,
 					(dma_addr_t) be64_to_cpu(data->addr),
 					 skb_frag_size(frag), PCI_DMA_TODEVICE);
 				++data;
@@ -745,7 +744,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 		/* Map fragments */
 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
 			frag = &skb_shinfo(skb)->frags[i];
-			dma = skb_frag_dma_map(&mdev->dev->pdev->dev, frag,
+			dma = skb_frag_dma_map(priv->ddev, frag,
 					       0, skb_frag_size(frag),
 					       DMA_TO_DEVICE);
 			data->addr = cpu_to_be64(dma);
@@ -757,7 +756,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 		/* Map linear part */
 		if (tx_info->linear) {
-			dma = pci_map_single(mdev->dev->pdev, skb->data + lso_header_size,
+			dma = dma_map_single(priv->ddev, skb->data + lso_header_size,
 					     skb_headlen(skb) - lso_header_size, PCI_DMA_TODEVICE);
 			data->addr = cpu_to_be64(dma);
 			data->lkey = cpu_to_be32(mdev->mr.key);
drivers/net/ethernet/mellanox/mlx4/mlx4_en.h +1 -0
index 174dc38..135200e 100644
@@ -482,6 +482,7 @@ struct mlx4_en_priv {
 	struct mlx4_en_stat_out_mbox hw_stats;
 	int vids[128];
 	bool wol;
+	struct device *ddev;
 };
 
 enum mlx4_en_wol {
Signature · 1 line
1.5.4.3
maintainer
#1 lore raw

From: Yevgeny Petrilin <[email protected]>
Date: Thu, 23 Feb 2012 15:34:44 +0200

Quoted 6 lines

Localized the pdev->dev, and using dma_map instead of pci_map There are multiple map/unmap operations on data path, optimizing those by saving redundant pointer access.

I doubt you can measure any improvement from this at all, all of the real cost is going to be in programming the hardware (both the networking chip and potentially any IOMMU hardware to setup the DMA mappings). A pointer deref in of your datastructures will be unnoticable.

When you make changes like this I want you to justify them with real data and facts.

I think this entire patch set was in very poor taste and judgment.

#2 lore raw
Quoted 11 lines
Quoted 6 lines

Localized the pdev->dev, and using dma_map instead of pci_map There are multiple map/unmap operations on data path, optimizing those by saving redundant pointer access.

I doubt you can measure any improvement from this at all, all of the real cost is going to be in programming the hardware (both the networking chip and potentially any IOMMU hardware to setup the DMA mappings). A pointer deref in of your datastructures will be unnoticable.

We actually did measure improvements with this changes.
Those places were found as hot spots when we did kernel profiling on some benchmarks. Making those changes actually helped us reduce cpu utilization in several %, and in some cases was the difference between being CPU bounded or reaching wire speed.
One scenario that we had improvement in when making this change is measuring Packet Rate for small (64 Byte) packets.

Quoted 3 lines

When you make changes like this I want you to justify them with real data and facts.

I agree, the explanation should have been part of the patch description.

Quoted 1 line

I think this entire patch set was in very poor taste and judgment.

Dave,
Thank you for the comprehensive review, please allow me to kindly disagree with the last statement.

The patch set is a result of performance work we did, and those changes helped us to achieve better results without hurting us in other areas.
I do understand you concern regarding patch 2/3, although we did test it in various scenarios without running into the issues you described, and there are other vendors doing the same.

I still think that patches 1 and 3 are good for our device,
I'll be happy to make changes in them if you find it necessary, and resubmit.

Thanks,
Yevgeny

#3 lore raw
Quoted 2 lines · Yevgeny Petrilin

The patch set is a result of performance work we did, and those changes helped us to achieve better results without hurting us in other areas.

Hi Dave,

Following Yevgeny's clarification, can this patch (3/3) go to net-next? as for patch 1/3, we
understand that the first thing to do in that area is add BQL support to mlx4_en, makes sense.

Or.

maintainer
#4 lore raw

From: Or Gerlitz <[email protected]>
Date: Thu, 1 Mar 2012 15:03:02 +0200

Quoted 2 lines

Following Yevgeny's clarification, can this patch (3/3) go to net-next?

Sure, please resubmit it.