[PATCH net-next 00/14] net/mlx4_en: Optimizations to TX flow

[PATCH net-next 00/14] net/mlx4_en: Optimizations to TX flow

[Amir Vadai]

Hi,

This patchset contains optimizations to TX flow in mlx4_en driver. It also introduce
setting/getting tx copybreak, to enable controlling inline threshold dynamically.

TX flow optimizations was authored and posted to the mailing list by Eric
Dumazet [1] as a single patch. I splitted this patch to smaller patches,
Reviewed it and tested.
Changed from original patch:
- s/iowrite32be/iowrite32/, since ring->doorbell_qpn is stored as be32

The tx copybreak patch was also suggested by Eric Dumazet, and was edited and
reviewed by me. User space patch will be sent after kernel code is ready.

I am sending this patchset now since the merge window is near and don't want to
miss it.

More work need to do:
- Disable BF when xmit_more is in use
- Make TSO use xmit_more too. Maybe by splitting small TSO packets in the
  driver itself, to avoid extra cpu/memory costs of GSO before the driver
- Fix mlx4_en_xmit buggy handling of queue full in the middle of a burst
  partially posted to send queue using xmit_more

Eric, I edited the patches to have you as the Author and the first
signed-off-by. I hope it is ok with you (I wasn't sure if it is ok to sign by
you), anyway all the credit to those changes should go to you.

Patchset was tested and applied over commit 1e203c1 "(net: sched:
suspicious RCU usage in qdisc_watchdog")

[1] - https://patchwork.ozlabs.org/patch/394256/

Thanks,
Amir

Amir Vadai (13):
  net/mlx4_en: Code cleanups in tx path
  net/mlx4_en: Align tx path structures to cache lines
  net/mlx4_en: Avoid calling bswap in tx fast path
  net/mlx4_en: tx_info allocated with kmalloc() instead of vmalloc()
  net/mlx4_en: Avoid a cache line miss in TX completion for single frag
    skb's
  net/mlx4_en: Use prefetch in tx path
  net/mlx4_en: Avoid false sharing in mlx4_en_en_process_tx_cq()
  net/mlx4_en: mlx4_en_xmit() reads ring->cons once, and ahead of time
    to avoid stalls
  net/mlx4_en: Use local var in tx flow for skb_shinfo(skb)
  net/mlx4_en: Use local var for skb_headlen(skb)
  net/mlx4_en: tx_info->ts_requested was not cleared
  net/mlx4_en: Enable the compiler to make is_inline() inlined
  ethtool: Ethtool parameter to dynamically change tx_copybreak

Eric Dumazet (1):
  net/mlx4_en: Use the new tx_copybreak to set inline threshold

 drivers/net/ethernet/mellanox/mlx4/en_ethtool.c |  44 ++++
 drivers/net/ethernet/mellanox/mlx4/en_tx.c      | 330 ++++++++++++++----------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h    |  90 ++++---
 include/linux/mlx4/device.h                     |   2 +-
 include/uapi/linux/ethtool.h                    |   1 +
 net/core/ethtool.c                              |   1 +
 6 files changed, 290 insertions(+), 178 deletions(-)

-- 
1.8.3.4

[PATCH net-next 01/14] net/mlx4_en: Code cleanups in tx path

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

- Remove unused variable ring->poll_cnt
- No need to set some fields if using blueflame
- Add missing const's
- Use unlikely
- Remove unneeded new line
- Make some comments more precise
- struct mlx4_bf @offset field reduced to unsigned int to save space

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 49 +++++++++++++++-------------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  1 -
 include/linux/mlx4/device.h                  |  2 +-
 3 files changed, 27 insertions(+), 25 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 0c50125..eaf23eb 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -191,7 +191,6 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
 	ring->prod = 0;
 	ring->cons = 0xffffffff;
 	ring->last_nr_txbb = 1;
-	ring->poll_cnt = 0;
 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
 	memset(ring->buf, 0, ring->buf_size);
 
@@ -512,7 +511,8 @@ static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 	return ring->buf + index * TXBB_SIZE;
 }
 
-static int is_inline(int inline_thold, struct sk_buff *skb, void **pfrag)
+static bool is_inline(int inline_thold, const struct sk_buff *skb,
+		      void **pfrag)
 {
 	void *ptr;
 
@@ -535,7 +535,7 @@ static int is_inline(int inline_thold, struct sk_buff *skb, void **pfrag)
 	return 0;
 }
 
-static int inline_size(struct sk_buff *skb)
+static int inline_size(const struct sk_buff *skb)
 {
 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
 	    <= MLX4_INLINE_ALIGN)
@@ -546,7 +546,8 @@ static int inline_size(struct sk_buff *skb)
 			     sizeof(struct mlx4_wqe_inline_seg), 16);
 }
 
-static int get_real_size(struct sk_buff *skb, struct net_device *dev,
+static int get_real_size(const struct sk_buff *skb,
+			 struct net_device *dev,
 			 int *lso_header_size)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
@@ -581,8 +582,10 @@ static int get_real_size(struct sk_buff *skb, struct net_device *dev,
 	return real_size;
 }
 
-static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
-			     int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
+static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
+			     const struct sk_buff *skb,
+			     int real_size, u16 *vlan_tag,
+			     int tx_ind, void *fragptr)
 {
 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
@@ -642,7 +645,8 @@ u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
 	return fallback(dev, skb) % rings_p_up + up * rings_p_up;
 }
 
-static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
+static void mlx4_bf_copy(void __iomem *dst, const void *src,
+			 unsigned int bytecnt)
 {
 	__iowrite64_copy(dst, src, bytecnt / 8);
 }
@@ -736,11 +740,10 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	tx_info->skb = skb;
 	tx_info->nr_txbb = nr_txbb;
 
+	data = &tx_desc->data;
 	if (lso_header_size)
 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
 						      DS_SIZE));
-	else
-		data = &tx_desc->data;
 
 	/* valid only for none inline segments */
 	tx_info->data_offset = (void *)data - (void *)tx_desc;
@@ -753,9 +756,9 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	if (is_inline(ring->inline_thold, skb, &fragptr)) {
 		tx_info->inl = 1;
 	} else {
-		/* Map fragments */
+		/* Map fragments if any */
 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
-			struct skb_frag_struct *frag;
+			const struct skb_frag_struct *frag;
 			dma_addr_t dma;
 
 			frag = &skb_shinfo(skb)->frags[i];
@@ -772,7 +775,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 			--data;
 		}
 
-		/* Map linear part */
+		/* Map linear part if needed */
 		if (tx_info->linear) {
 			u32 byte_count = skb_headlen(skb) - lso_header_size;
 			dma_addr_t dma;
@@ -795,18 +798,14 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	 * For timestamping add flag to skb_shinfo and
 	 * set flag for further reference
 	 */
-	if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
-	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
-		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
+	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
+		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
+		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
 		tx_info->ts_requested = 1;
 	}
 
 	/* Prepare ctrl segement apart opcode+ownership, which depends on
 	 * whether LSO is used */
-	tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
-	tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
-		!!vlan_tx_tag_present(skb);
-	tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
@@ -852,7 +851,6 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
 		ring->packets++;
-
 	}
 	ring->bytes += tx_info->nr_bytes;
 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
@@ -874,7 +872,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	ring->prod += nr_txbb;
 
 	/* If we used a bounce buffer then copy descriptor back into place */
-	if (bounce)
+	if (unlikely(bounce))
 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
 
 	skb_tx_timestamp(skb);
@@ -894,13 +892,18 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 		wmb();
 
-		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
-		     desc_size);
+		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
+			     desc_size);
 
 		wmb();
 
 		ring->bf.offset ^= ring->bf.buf_size;
 	} else {
+		tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
+		tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
+			!!vlan_tx_tag_present(skb);
+		tx_desc->ctrl.fence_size = real_size;
+
 		/* Ensure new descriptor hits memory
 		 * before setting ownership of this descriptor to HW
 		 */
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 84c9d5d..e54b653 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -263,7 +263,6 @@ struct mlx4_en_tx_ring {
 	u32 buf_size;
 	u32 doorbell_qpn;
 	void *buf;
-	u16 poll_cnt;
 	struct mlx4_en_tx_info *tx_info;
 	u8 *bounce_buf;
 	u8 queue_index;
diff --git a/include/linux/mlx4/device.h b/include/linux/mlx4/device.h
index b2f8ab9..37e4404 100644
--- a/include/linux/mlx4/device.h
+++ b/include/linux/mlx4/device.h
@@ -583,7 +583,7 @@ struct mlx4_uar {
 };
 
 struct mlx4_bf {
-	unsigned long		offset;
+	unsigned int		offset;
 	int			buf_size;
 	struct mlx4_uar	       *uar;
 	void __iomem	       *reg;
-- 
1.8.3.4

[PATCH net-next 02/14] net/mlx4_en: Align tx path structures to cache lines

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Reorganize struct mlx4_en_tx_ring to have:
- One cache line containing last_nr_txbb & cons & wake_queue, used by tx
  completion.
- One cache line containing fields dirtied by mlx4_en_xmit()
- Following part is read mostly and shared by cpus.

Align struct mlx4_en_tx_info to a cache line

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 86 +++++++++++++++-------------
 1 file changed, 46 insertions(+), 40 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index e54b653..b7bde95 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -216,13 +216,13 @@ enum cq_type {
 
 struct mlx4_en_tx_info {
 	struct sk_buff *skb;
-	u32 nr_txbb;
-	u32 nr_bytes;
-	u8 linear;
-	u8 data_offset;
-	u8 inl;
-	u8 ts_requested;
-};
+	u32		nr_txbb;
+	u32		nr_bytes;
+	u8		linear;
+	u8		data_offset;
+	u8		inl;
+	u8		ts_requested;
+} ____cacheline_aligned_in_smp;
 
 
 #define MLX4_EN_BIT_DESC_OWN	0x80000000
@@ -253,40 +253,46 @@ struct mlx4_en_rx_alloc {
 };
 
 struct mlx4_en_tx_ring {
+	/* cache line used and dirtied in tx completion
+	 * (mlx4_en_free_tx_buf())
+	 */
+	u32			last_nr_txbb;
+	u32			cons;
+	unsigned long		wake_queue;
+
+	/* cache line used and dirtied in mlx4_en_xmit() */
+	u32			prod ____cacheline_aligned_in_smp;
+	unsigned long		bytes;
+	unsigned long		packets;
+	unsigned long		tx_csum;
+	unsigned long		tso_packets;
+	unsigned long		xmit_more;
+	struct mlx4_bf		bf;
+	unsigned long		queue_stopped;
+
+	/* Following part should be mostly read */
+	cpumask_t		affinity_mask;
+	struct mlx4_qp		qp;
 	struct mlx4_hwq_resources wqres;
-	u32 size ; /* number of TXBBs */
-	u32 size_mask;
-	u16 stride;
-	u16 cqn;	/* index of port CQ associated with this ring */
-	u32 prod;
-	u32 cons;
-	u32 buf_size;
-	u32 doorbell_qpn;
-	void *buf;
-	struct mlx4_en_tx_info *tx_info;
-	u8 *bounce_buf;
-	u8 queue_index;
-	cpumask_t affinity_mask;
-	u32 last_nr_txbb;
-	struct mlx4_qp qp;
-	struct mlx4_qp_context context;
-	int qpn;
-	enum mlx4_qp_state qp_state;
-	struct mlx4_srq dummy;
-	unsigned long bytes;
-	unsigned long packets;
-	unsigned long tx_csum;
-	unsigned long queue_stopped;
-	unsigned long wake_queue;
-	unsigned long tso_packets;
-	unsigned long xmit_more;
-	struct mlx4_bf bf;
-	bool bf_enabled;
-	bool bf_alloced;
-	struct netdev_queue *tx_queue;
-	int hwtstamp_tx_type;
-	int inline_thold;
-};
+	u32			size; /* number of TXBBs */
+	u32			size_mask;
+	u16			stride;
+	u16			cqn;	/* index of port CQ associated with this ring */
+	u32			buf_size;
+	u32			doorbell_qpn;
+	void			*buf;
+	struct mlx4_en_tx_info	*tx_info;
+	u8			*bounce_buf;
+	struct mlx4_qp_context	context;
+	int			qpn;
+	enum mlx4_qp_state	qp_state;
+	u8			queue_index;
+	bool			bf_enabled;
+	bool			bf_alloced;
+	struct netdev_queue	*tx_queue;
+	int			hwtstamp_tx_type;
+	int			inline_thold;
+} ____cacheline_aligned_in_smp;
 
 struct mlx4_en_rx_desc {
 	/* actual number of entries depends on rx ring stride */
-- 
1.8.3.4

[PATCH net-next 03/14] net/mlx4_en: Avoid calling bswap in tx fast path

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

- doorbell_qpn is stored in the cpu_to_be32() way to avoid bswap() in fast
  path.
- mdev->mr.key stored in ring->mr_key to also avoid bswap() and access to
  cold cache line.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 17 ++++++++++-------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  3 ++-
 2 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index eaf23eb..322cda3 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -195,7 +195,8 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
 	memset(ring->buf, 0, ring->buf_size);
 
 	ring->qp_state = MLX4_QP_STATE_RST;
-	ring->doorbell_qpn = ring->qp.qpn << 8;
+	ring->doorbell_qpn = cpu_to_be32(ring->qp.qpn << 8);
+	ring->mr_key = cpu_to_be32(mdev->mr.key);
 
 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
 				ring->cqn, user_prio, &ring->context);
@@ -654,7 +655,6 @@ static void mlx4_bf_copy(void __iomem *dst, const void *src,
 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
-	struct mlx4_en_dev *mdev = priv->mdev;
 	struct device *ddev = priv->ddev;
 	struct mlx4_en_tx_ring *ring;
 	struct mlx4_en_tx_desc *tx_desc;
@@ -769,7 +769,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 				goto tx_drop_unmap;
 
 			data->addr = cpu_to_be64(dma);
-			data->lkey = cpu_to_be32(mdev->mr.key);
+			data->lkey = ring->mr_key;
 			wmb();
 			data->byte_count = cpu_to_be32(skb_frag_size(frag));
 			--data;
@@ -787,7 +787,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 				goto tx_drop_unmap;
 
 			data->addr = cpu_to_be64(dma);
-			data->lkey = cpu_to_be32(mdev->mr.key);
+			data->lkey = ring->mr_key;
 			wmb();
 			data->byte_count = cpu_to_be32(byte_count);
 		}
@@ -879,9 +879,12 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
 
+	real_size = (real_size / 16) & 0x3f;
+
 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
 	    !vlan_tx_tag_present(skb) && send_doorbell) {
-		tx_desc->ctrl.bf_qpn |= cpu_to_be32(ring->doorbell_qpn);
+		tx_desc->ctrl.bf_qpn = ring->doorbell_qpn |
+				       cpu_to_be32(real_size);
 
 		op_own |= htonl((bf_index & 0xffff) << 8);
 		/* Ensure new descriptor hits memory
@@ -911,8 +914,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 		tx_desc->ctrl.owner_opcode = op_own;
 		if (send_doorbell) {
 			wmb();
-			iowrite32be(ring->doorbell_qpn,
-				    ring->bf.uar->map + MLX4_SEND_DOORBELL);
+			iowrite32(ring->doorbell_qpn,
+				  ring->bf.uar->map + MLX4_SEND_DOORBELL);
 		} else {
 			ring->xmit_more++;
 		}
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index b7bde95..ab34461 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -279,7 +279,8 @@ struct mlx4_en_tx_ring {
 	u16			stride;
 	u16			cqn;	/* index of port CQ associated with this ring */
 	u32			buf_size;
-	u32			doorbell_qpn;
+	__be32			doorbell_qpn;
+	__be32			mr_key;
 	void			*buf;
 	struct mlx4_en_tx_info	*tx_info;
 	u8			*bounce_buf;
-- 
1.8.3.4

[PATCH net-next 04/14] net/mlx4_en: tx_info allocated with kmalloc() instead of vmalloc()

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Try to allocate using kmalloc_node() first, only on failure use
vmalloc()

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 322cda3..1447906 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -68,7 +68,7 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
 	ring->inline_thold = priv->prof->inline_thold;
 
 	tmp = size * sizeof(struct mlx4_en_tx_info);
-	ring->tx_info = vmalloc_node(tmp, node);
+	ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
 	if (!ring->tx_info) {
 		ring->tx_info = vmalloc(tmp);
 		if (!ring->tx_info) {
@@ -151,7 +151,7 @@ err_bounce:
 	kfree(ring->bounce_buf);
 	ring->bounce_buf = NULL;
 err_info:
-	vfree(ring->tx_info);
+	kvfree(ring->tx_info);
 	ring->tx_info = NULL;
 err_ring:
 	kfree(ring);
@@ -174,7 +174,7 @@ void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
 	kfree(ring->bounce_buf);
 	ring->bounce_buf = NULL;
-	vfree(ring->tx_info);
+	kvfree(ring->tx_info);
 	ring->tx_info = NULL;
 	kfree(ring);
 	*pring = NULL;
-- 
1.8.3.4

[PATCH net-next 05/14] net/mlx4_en: Avoid a cache line miss in TX completion for single frag skb's

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Add frag0_dma/frag0_byte_count into mlx4_en_tx_info to avoid a cache
line miss in TX completion for frames having one dma element.  (We avoid
reading back the tx descriptor)

Note this could be extended to 2/3 dma elements later, as we have free
room in mlx4_en_tx_info

Also, mlx4_en_free_tx_desc() no longer accesses skb_shinfo(). We use a
new nr_maps fields in mlx4_en_tx_info to avoid 2 or 3 cache misses.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c   | 83 +++++++++++++++-------------
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h |  3 +
 2 files changed, 49 insertions(+), 37 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 1447906..edc4a88 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -259,38 +259,40 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 				struct mlx4_en_tx_ring *ring,
 				int index, u8 owner, u64 timestamp)
 {
-	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;
-	struct sk_buff *skb = tx_info->skb;
-	struct skb_frag_struct *frag;
 	void *end = ring->buf + ring->buf_size;
-	int frags = skb_shinfo(skb)->nr_frags;
+	struct sk_buff *skb = tx_info->skb;
+	int nr_maps = tx_info->nr_maps;
 	int i;
-	struct skb_shared_hwtstamps hwts;
 
-	if (timestamp) {
-		mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
+	if (unlikely(timestamp)) {
+		struct skb_shared_hwtstamps hwts;
+
+		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
 		skb_tstamp_tx(skb, &hwts);
 	}
 
 	/* Optimize the common case when there are no wraparounds */
 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
 		if (!tx_info->inl) {
-			if (tx_info->linear) {
+			if (tx_info->linear)
 				dma_unmap_single(priv->ddev,
-					(dma_addr_t) be64_to_cpu(data->addr),
-					 be32_to_cpu(data->byte_count),
-					 PCI_DMA_TODEVICE);
-				++data;
-			}
-
-			for (i = 0; i < frags; i++) {
-				frag = &skb_shinfo(skb)->frags[i];
+						tx_info->map0_dma,
+						tx_info->map0_byte_count,
+						PCI_DMA_TODEVICE);
+			else
+				dma_unmap_page(priv->ddev,
+					       tx_info->map0_dma,
+					       tx_info->map0_byte_count,
+					       PCI_DMA_TODEVICE);
+			for (i = 1; i < nr_maps; i++) {
+				data++;
 				dma_unmap_page(priv->ddev,
-					(dma_addr_t) be64_to_cpu(data[i].addr),
-					skb_frag_size(frag), PCI_DMA_TODEVICE);
+					(dma_addr_t)be64_to_cpu(data->addr),
+					be32_to_cpu(data->byte_count),
+					PCI_DMA_TODEVICE);
 			}
 		}
 	} else {
@@ -299,23 +301,25 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 				data = ring->buf + ((void *)data - end);
 			}
 
-			if (tx_info->linear) {
+			if (tx_info->linear)
 				dma_unmap_single(priv->ddev,
-					(dma_addr_t) be64_to_cpu(data->addr),
-					 be32_to_cpu(data->byte_count),
-					 PCI_DMA_TODEVICE);
-				++data;
-			}
-
-			for (i = 0; i < frags; i++) {
+						tx_info->map0_dma,
+						tx_info->map0_byte_count,
+						PCI_DMA_TODEVICE);
+			else
+				dma_unmap_page(priv->ddev,
+					       tx_info->map0_dma,
+					       tx_info->map0_byte_count,
+					       PCI_DMA_TODEVICE);
+			for (i = 1; i < nr_maps; i++) {
+				data++;
 				/* Check for wraparound before unmapping */
 				if ((void *) data >= end)
 					data = ring->buf;
-				frag = &skb_shinfo(skb)->frags[i];
 				dma_unmap_page(priv->ddev,
-					(dma_addr_t) be64_to_cpu(data->addr),
-					 skb_frag_size(frag), PCI_DMA_TODEVICE);
-				++data;
+					(dma_addr_t)be64_to_cpu(data->addr),
+					be32_to_cpu(data->byte_count),
+					PCI_DMA_TODEVICE);
 			}
 		}
 	}
@@ -751,19 +755,22 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
 			   !is_inline(ring->inline_thold, skb, NULL)) ? 1 : 0;
 
-	data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
+	tx_info->nr_maps = skb_shinfo(skb)->nr_frags + tx_info->linear;
+	data += tx_info->nr_maps - 1;
 
 	if (is_inline(ring->inline_thold, skb, &fragptr)) {
 		tx_info->inl = 1;
 	} else {
+		dma_addr_t dma = 0;
+		u32 byte_count = 0;
+
 		/* Map fragments if any */
 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
 			const struct skb_frag_struct *frag;
-			dma_addr_t dma;
-
 			frag = &skb_shinfo(skb)->frags[i];
+			byte_count = skb_frag_size(frag);
 			dma = skb_frag_dma_map(ddev, frag,
-					       0, skb_frag_size(frag),
+					       0, byte_count,
 					       DMA_TO_DEVICE);
 			if (dma_mapping_error(ddev, dma))
 				goto tx_drop_unmap;
@@ -771,14 +778,13 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 			data->addr = cpu_to_be64(dma);
 			data->lkey = ring->mr_key;
 			wmb();
-			data->byte_count = cpu_to_be32(skb_frag_size(frag));
+			data->byte_count = cpu_to_be32(byte_count);
 			--data;
 		}
 
 		/* Map linear part if needed */
 		if (tx_info->linear) {
-			u32 byte_count = skb_headlen(skb) - lso_header_size;
-			dma_addr_t dma;
+			byte_count = skb_headlen(skb) - lso_header_size;
 
 			dma = dma_map_single(ddev, skb->data +
 					     lso_header_size, byte_count,
@@ -792,6 +798,9 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 			data->byte_count = cpu_to_be32(byte_count);
 		}
 		tx_info->inl = 0;
+		/* tx completion can avoid cache line miss for common cases */
+		tx_info->map0_dma = dma;
+		tx_info->map0_byte_count = byte_count;
 	}
 
 	/*
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index ab34461..a904030 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -216,12 +216,15 @@ enum cq_type {
 
 struct mlx4_en_tx_info {
 	struct sk_buff *skb;
+	dma_addr_t	map0_dma;
+	u32		map0_byte_count;
 	u32		nr_txbb;
 	u32		nr_bytes;
 	u8		linear;
 	u8		data_offset;
 	u8		inl;
 	u8		ts_requested;
+	u8		nr_maps;
 } ____cacheline_aligned_in_smp;
 
 
-- 
1.8.3.4

[PATCH net-next 06/14] net/mlx4_en: Use prefetch in tx path

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

mlx4_en_free_tx_desc() uses a prefetchw(&skb->users) to speed up
consume_skb()
prefetchw(&ring->tx_queue->dql) to speed up BQL update

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index edc4a88..b96627c 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -37,6 +37,7 @@
 #include <linux/mlx4/qp.h>
 #include <linux/skbuff.h>
 #include <linux/if_vlan.h>
+#include <linux/prefetch.h>
 #include <linux/vmalloc.h>
 #include <linux/tcp.h>
 #include <linux/ip.h>
@@ -267,6 +268,11 @@ static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 	int nr_maps = tx_info->nr_maps;
 	int i;
 
+	/* We do not touch skb here, so prefetch skb->users location
+	 * to speedup consume_skb()
+	 */
+	prefetchw(&skb->users);
+
 	if (unlikely(timestamp)) {
 		struct skb_shared_hwtstamps hwts;
 
@@ -385,6 +391,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	if (!priv->port_up)
 		return true;
 
+	prefetchw(&ring->tx_queue->dql.limit);
 	index = cons_index & size_mask;
 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
 	ring_index = ring->cons & size_mask;
@@ -722,6 +729,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 		}
 	}
 
+	prefetchw(&ring->tx_queue->dql);
+
 	/* Track current inflight packets for performance analysis */
 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
 			 (u32) (ring->prod - ring->cons - 1));
-- 
1.8.3.4

[PATCH net-next 07/14] net/mlx4_en: Avoid false sharing in mlx4_en_en_process_tx_cq()

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

mlx4_en_process_tx_cq() carefully fetches and writes ring->last_nr_txbb
and ring->cons only one time to avoid false sharing

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index b96627c..d9aaeb2 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -387,6 +387,8 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	u64 timestamp = 0;
 	int done = 0;
 	int budget = priv->tx_work_limit;
+	u32 last_nr_txbb;
+	u32 ring_cons;
 
 	if (!priv->port_up)
 		return true;
@@ -394,7 +396,9 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	prefetchw(&ring->tx_queue->dql.limit);
 	index = cons_index & size_mask;
 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
-	ring_index = ring->cons & size_mask;
+	last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
+	ring_cons = ACCESS_ONCE(ring->cons);
+	ring_index = ring_cons & size_mask;
 	stamp_index = ring_index;
 
 	/* Process all completed CQEs */
@@ -419,19 +423,19 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
 
 		do {
-			txbbs_skipped += ring->last_nr_txbb;
-			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
+			txbbs_skipped += last_nr_txbb;
+			ring_index = (ring_index + last_nr_txbb) & size_mask;
 			if (ring->tx_info[ring_index].ts_requested)
 				timestamp = mlx4_en_get_cqe_ts(cqe);
 
 			/* free next descriptor */
-			ring->last_nr_txbb = mlx4_en_free_tx_desc(
+			last_nr_txbb = mlx4_en_free_tx_desc(
 					priv, ring, ring_index,
-					!!((ring->cons + txbbs_skipped) &
+					!!((ring_cons + txbbs_skipped) &
 					ring->size), timestamp);
 
 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
-					  !!((ring->cons + txbbs_stamp) &
+					  !!((ring_cons + txbbs_stamp) &
 						ring->size));
 			stamp_index = ring_index;
 			txbbs_stamp = txbbs_skipped;
@@ -452,7 +456,11 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 	mcq->cons_index = cons_index;
 	mlx4_cq_set_ci(mcq);
 	wmb();
-	ring->cons += txbbs_skipped;
+
+	/* we want to dirty this cache line once */
+	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
+	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
+
 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
 
 	/*
-- 
1.8.3.4

[PATCH net-next 08/14] net/mlx4_en: mlx4_en_xmit() reads ring->cons once, and ahead of time to avoid stalls

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index d9aaeb2..99875c8 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -691,10 +691,17 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	void *fragptr;
 	bool bounce = false;
 	bool send_doorbell;
+	u32 ring_cons;
 
 	if (!priv->port_up)
 		goto tx_drop;
 
+	tx_ind = skb_get_queue_mapping(skb);
+	ring = priv->tx_ring[tx_ind];
+
+	/* fetch ring->cons far ahead before needing it to avoid stall */
+	ring_cons = ACCESS_ONCE(ring->cons);
+
 	real_size = get_real_size(skb, dev, &lso_header_size);
 	if (unlikely(!real_size))
 		goto tx_drop;
@@ -708,13 +715,11 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 		goto tx_drop;
 	}
 
-	tx_ind = skb->queue_mapping;
-	ring = priv->tx_ring[tx_ind];
 	if (vlan_tx_tag_present(skb))
 		vlan_tag = vlan_tx_tag_get(skb);
 
 	/* Check available TXBBs And 2K spare for prefetch */
-	if (unlikely(((int)(ring->prod - ring->cons)) >
+	if (unlikely(((int)(ring->prod - ring_cons)) >
 		     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
 		/* every full Tx ring stops queue */
 		netif_tx_stop_queue(ring->tx_queue);
@@ -728,7 +733,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 		 */
 		wmb();
 
-		if (unlikely(((int)(ring->prod - ring->cons)) <=
+		ring_cons = ACCESS_ONCE(ring->cons);
+		if (unlikely(((int)(ring->prod - ring_cons)) <=
 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
 			netif_tx_wake_queue(ring->tx_queue);
 			ring->wake_queue++;
@@ -741,7 +747,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	/* Track current inflight packets for performance analysis */
 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
-			 (u32) (ring->prod - ring->cons - 1));
+			 (u32)(ring->prod - ring_cons - 1));
 
 	/* Packet is good - grab an index and transmit it */
 	index = ring->prod & ring->size_mask;
-- 
1.8.3.4

[PATCH net-next 09/14] net/mlx4_en: Use local var in tx flow for skb_shinfo(skb)

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Acces skb_shinfo(skb) once in tx flow.
Also, rename @i variable to @i_frag to avoid confusion, as the "goto
tx_drop_unmap;" relied on this @i variable.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 58 +++++++++++++++++-------------
 1 file changed, 34 insertions(+), 24 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 99875c8..aa05b09 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -532,13 +532,14 @@ static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 }
 
 static bool is_inline(int inline_thold, const struct sk_buff *skb,
+		      const struct skb_shared_info *shinfo,
 		      void **pfrag)
 {
 	void *ptr;
 
 	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
-		if (skb_shinfo(skb)->nr_frags == 1) {
-			ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
+		if (shinfo->nr_frags == 1) {
+			ptr = skb_frag_address_safe(&shinfo->frags[0]);
 			if (unlikely(!ptr))
 				return 0;
 
@@ -546,7 +547,7 @@ static bool is_inline(int inline_thold, const struct sk_buff *skb,
 				*pfrag = ptr;
 
 			return 1;
-		} else if (unlikely(skb_shinfo(skb)->nr_frags))
+		} else if (unlikely(shinfo->nr_frags))
 			return 0;
 		else
 			return 1;
@@ -567,18 +568,19 @@ static int inline_size(const struct sk_buff *skb)
 }
 
 static int get_real_size(const struct sk_buff *skb,
+			 const struct skb_shared_info *shinfo,
 			 struct net_device *dev,
 			 int *lso_header_size)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	int real_size;
 
-	if (skb_is_gso(skb)) {
+	if (shinfo->gso_size) {
 		if (skb->encapsulation)
 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
 		else
 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
-		real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
+		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
 			ALIGN(*lso_header_size + 4, DS_SIZE);
 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
 			/* We add a segment for the skb linear buffer only if
@@ -593,8 +595,8 @@ static int get_real_size(const struct sk_buff *skb,
 		}
 	} else {
 		*lso_header_size = 0;
-		if (!is_inline(priv->prof->inline_thold, skb, NULL))
-			real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
+		if (!is_inline(priv->prof->inline_thold, skb, shinfo, NULL))
+			real_size = CTRL_SIZE + (shinfo->nr_frags + 1) * DS_SIZE;
 		else
 			real_size = inline_size(skb);
 	}
@@ -604,6 +606,7 @@ static int get_real_size(const struct sk_buff *skb,
 
 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 			     const struct sk_buff *skb,
+			     const struct skb_shared_info *shinfo,
 			     int real_size, u16 *vlan_tag,
 			     int tx_ind, void *fragptr)
 {
@@ -619,9 +622,9 @@ static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 			       MIN_PKT_LEN - skb->len);
 		}
 		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
-		if (skb_shinfo(skb)->nr_frags)
+		if (shinfo->nr_frags)
 			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
-			       skb_frag_size(&skb_shinfo(skb)->frags[0]));
+			       skb_frag_size(&shinfo->frags[0]));
 
 	} else {
 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
@@ -639,9 +642,10 @@ static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 			inl = (void *) (inl + 1) + spc;
 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
 					skb_headlen(skb) - spc);
-			if (skb_shinfo(skb)->nr_frags)
+			if (shinfo->nr_frags)
 				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
-					fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
+				       fragptr,
+				       skb_frag_size(&shinfo->frags[0]));
 		}
 
 		wmb();
@@ -673,6 +677,7 @@ static void mlx4_bf_copy(void __iomem *dst, const void *src,
 
 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 {
+	struct skb_shared_info *shinfo = skb_shinfo(skb);
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	struct device *ddev = priv->ddev;
 	struct mlx4_en_tx_ring *ring;
@@ -686,7 +691,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	u32 index, bf_index;
 	__be32 op_own;
 	u16 vlan_tag = 0;
-	int i;
+	int i_frag;
 	int lso_header_size;
 	void *fragptr;
 	bool bounce = false;
@@ -702,7 +707,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	/* fetch ring->cons far ahead before needing it to avoid stall */
 	ring_cons = ACCESS_ONCE(ring->cons);
 
-	real_size = get_real_size(skb, dev, &lso_header_size);
+	real_size = get_real_size(skb, shinfo, dev, &lso_header_size);
 	if (unlikely(!real_size))
 		goto tx_drop;
 
@@ -776,21 +781,22 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	tx_info->data_offset = (void *)data - (void *)tx_desc;
 
 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
-			   !is_inline(ring->inline_thold, skb, NULL)) ? 1 : 0;
+			   !is_inline(ring->inline_thold, skb, shinfo, NULL)) ? 1 : 0;
 
-	tx_info->nr_maps = skb_shinfo(skb)->nr_frags + tx_info->linear;
+	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
 	data += tx_info->nr_maps - 1;
 
-	if (is_inline(ring->inline_thold, skb, &fragptr)) {
+	if (is_inline(ring->inline_thold, skb, shinfo, &fragptr)) {
 		tx_info->inl = 1;
 	} else {
 		dma_addr_t dma = 0;
 		u32 byte_count = 0;
 
 		/* Map fragments if any */
-		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
+		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
 			const struct skb_frag_struct *frag;
-			frag = &skb_shinfo(skb)->frags[i];
+
+			frag = &shinfo->frags[i_frag];
 			byte_count = skb_frag_size(frag);
 			dma = skb_frag_dma_map(ddev, frag,
 					       0, byte_count,
@@ -858,6 +864,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	/* Handle LSO (TSO) packets */
 	if (lso_header_size) {
+		int i;
+
 		/* Mark opcode as LSO */
 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
 			((ring->prod & ring->size) ?
@@ -865,15 +873,16 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
 		/* Fill in the LSO prefix */
 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
-			skb_shinfo(skb)->gso_size << 16 | lso_header_size);
+			shinfo->gso_size << 16 | lso_header_size);
 
 		/* Copy headers;
 		 * note that we already verified that it is linear */
 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
 
 		ring->tso_packets++;
-		i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
-			!!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
+
+		i = ((skb->len - lso_header_size) / shinfo->gso_size) +
+			!!((skb->len - lso_header_size) % shinfo->gso_size);
 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
 		ring->packets += i;
 	} else {
@@ -889,7 +898,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
 
 	if (tx_info->inl) {
-		build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
+		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
+				 tx_ind, fragptr);
 		tx_info->inl = 1;
 	}
 
@@ -958,8 +968,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 tx_drop_unmap:
 	en_err(priv, "DMA mapping error\n");
 
-	for (i++; i < skb_shinfo(skb)->nr_frags; i++) {
-		data++;
+	while (++i_frag < shinfo->nr_frags) {
+		++data;
 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
 			       be32_to_cpu(data->byte_count),
 			       PCI_DMA_TODEVICE);
-- 
1.8.3.4

[PATCH net-next 10/14] net/mlx4_en: Use local var for skb_headlen(skb)

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Access skb_headlen() once in tx flow

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index aa05b09..e00841a 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -612,6 +612,7 @@ static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 {
 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
+	unsigned int hlen = skb_headlen(skb);
 
 	if (skb->len <= spc) {
 		if (likely(skb->len >= MIN_PKT_LEN)) {
@@ -621,19 +622,19 @@ static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 			memset(((void *)(inl + 1)) + skb->len, 0,
 			       MIN_PKT_LEN - skb->len);
 		}
-		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
+		skb_copy_from_linear_data(skb, inl + 1, hlen);
 		if (shinfo->nr_frags)
-			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
+			memcpy(((void *)(inl + 1)) + hlen, fragptr,
 			       skb_frag_size(&shinfo->frags[0]));
 
 	} else {
 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
-		if (skb_headlen(skb) <= spc) {
-			skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
-			if (skb_headlen(skb) < spc) {
-				memcpy(((void *)(inl + 1)) + skb_headlen(skb),
-					fragptr, spc - skb_headlen(skb));
-				fragptr +=  spc - skb_headlen(skb);
+		if (hlen <= spc) {
+			skb_copy_from_linear_data(skb, inl + 1, hlen);
+			if (hlen < spc) {
+				memcpy(((void *)(inl + 1)) + hlen,
+				       fragptr, spc - hlen);
+				fragptr +=  spc - hlen;
 			}
 			inl = (void *) (inl + 1) + spc;
 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
@@ -641,9 +642,9 @@ static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
 			skb_copy_from_linear_data(skb, inl + 1, spc);
 			inl = (void *) (inl + 1) + spc;
 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
-					skb_headlen(skb) - spc);
+							 hlen - spc);
 			if (shinfo->nr_frags)
-				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
+				memcpy(((void *)(inl + 1)) + hlen - spc,
 				       fragptr,
 				       skb_frag_size(&shinfo->frags[0]));
 		}
-- 
1.8.3.4

[PATCH net-next 11/14] net/mlx4_en: tx_info->ts_requested was not cleared

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Properly clear tx_info->ts_requested

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index e00841a..2c03b55 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -837,6 +837,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	 * For timestamping add flag to skb_shinfo and
 	 * set flag for further reference
 	 */
+	tx_info->ts_requested = 0;
 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
-- 
1.8.3.4

[PATCH net-next 12/14] net/mlx4_en: Enable the compiler to make is_inline() inlined

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Reorganize code to call is_inline() once, so compiler can inline it

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_tx.c | 67 +++++++++++++++++-------------
 1 file changed, 38 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 2c03b55..f0080c5 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -531,29 +531,32 @@ static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 	return ring->buf + index * TXBB_SIZE;
 }
 
+/* Decide if skb can be inlined in tx descriptor to avoid dma mapping
+ *
+ * It seems strange we do not simply use skb_copy_bits().
+ * This would allow to inline all skbs iff skb->len <= inline_thold
+ *
+ * Note that caller already checked skb was not a gso packet
+ */
 static bool is_inline(int inline_thold, const struct sk_buff *skb,
 		      const struct skb_shared_info *shinfo,
 		      void **pfrag)
 {
 	void *ptr;
 
-	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
-		if (shinfo->nr_frags == 1) {
-			ptr = skb_frag_address_safe(&shinfo->frags[0]);
-			if (unlikely(!ptr))
-				return 0;
-
-			if (pfrag)
-				*pfrag = ptr;
+	if (skb->len > inline_thold || !inline_thold)
+		return false;
 
-			return 1;
-		} else if (unlikely(shinfo->nr_frags))
-			return 0;
-		else
-			return 1;
+	if (shinfo->nr_frags == 1) {
+		ptr = skb_frag_address_safe(&shinfo->frags[0]);
+		if (unlikely(!ptr))
+			return false;
+		*pfrag = ptr;
+		return true;
 	}
-
-	return 0;
+	if (shinfo->nr_frags)
+		return false;
+	return true;
 }
 
 static int inline_size(const struct sk_buff *skb)
@@ -570,12 +573,15 @@ static int inline_size(const struct sk_buff *skb)
 static int get_real_size(const struct sk_buff *skb,
 			 const struct skb_shared_info *shinfo,
 			 struct net_device *dev,
-			 int *lso_header_size)
+			 int *lso_header_size,
+			 bool *inline_ok,
+			 void **pfrag)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 	int real_size;
 
 	if (shinfo->gso_size) {
+		*inline_ok = false;
 		if (skb->encapsulation)
 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
 		else
@@ -595,10 +601,14 @@ static int get_real_size(const struct sk_buff *skb,
 		}
 	} else {
 		*lso_header_size = 0;
-		if (!is_inline(priv->prof->inline_thold, skb, shinfo, NULL))
-			real_size = CTRL_SIZE + (shinfo->nr_frags + 1) * DS_SIZE;
-		else
+		*inline_ok = is_inline(priv->prof->inline_thold, skb,
+				       shinfo, pfrag);
+
+		if (*inline_ok)
 			real_size = inline_size(skb);
+		else
+			real_size = CTRL_SIZE +
+				    (shinfo->nr_frags + 1) * DS_SIZE;
 	}
 
 	return real_size;
@@ -694,9 +704,10 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	u16 vlan_tag = 0;
 	int i_frag;
 	int lso_header_size;
-	void *fragptr;
+	void *fragptr = NULL;
 	bool bounce = false;
 	bool send_doorbell;
+	bool inline_ok;
 	u32 ring_cons;
 
 	if (!priv->port_up)
@@ -708,7 +719,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	/* fetch ring->cons far ahead before needing it to avoid stall */
 	ring_cons = ACCESS_ONCE(ring->cons);
 
-	real_size = get_real_size(skb, shinfo, dev, &lso_header_size);
+	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
+				  &inline_ok, &fragptr);
 	if (unlikely(!real_size))
 		goto tx_drop;
 
@@ -781,15 +793,15 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	/* valid only for none inline segments */
 	tx_info->data_offset = (void *)data - (void *)tx_desc;
 
+	tx_info->inl = inline_ok;
+
 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
-			   !is_inline(ring->inline_thold, skb, shinfo, NULL)) ? 1 : 0;
+			   !inline_ok) ? 1 : 0;
 
 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
 	data += tx_info->nr_maps - 1;
 
-	if (is_inline(ring->inline_thold, skb, shinfo, &fragptr)) {
-		tx_info->inl = 1;
-	} else {
+	if (!tx_info->inl) {
 		dma_addr_t dma = 0;
 		u32 byte_count = 0;
 
@@ -827,7 +839,6 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 			wmb();
 			data->byte_count = cpu_to_be32(byte_count);
 		}
-		tx_info->inl = 0;
 		/* tx completion can avoid cache line miss for common cases */
 		tx_info->map0_dma = dma;
 		tx_info->map0_byte_count = byte_count;
@@ -899,11 +910,9 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
 
-	if (tx_info->inl) {
+	if (tx_info->inl)
 		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
 				 tx_ind, fragptr);
-		tx_info->inl = 1;
-	}
 
 	if (skb->encapsulation) {
 		struct iphdr *ipv4 = (struct iphdr *)skb_inner_network_header(skb);
-- 
1.8.3.4

[PATCH net-next 13/14] ethtool: Ethtool parameter to dynamically change tx_copybreak

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Use new ethtool [sg]et_tunable() to set tx_copybread (inline threshold)

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 include/uapi/linux/ethtool.h | 1 +
 net/core/ethtool.c           | 1 +
 2 files changed, 2 insertions(+)

diff --git a/include/uapi/linux/ethtool.h b/include/uapi/linux/ethtool.h
index 7a364f2..99b4305 100644
--- a/include/uapi/linux/ethtool.h
+++ b/include/uapi/linux/ethtool.h
@@ -212,6 +212,7 @@ struct ethtool_value {
 enum tunable_id {
 	ETHTOOL_ID_UNSPEC,
 	ETHTOOL_RX_COPYBREAK,
+	ETHTOOL_TX_COPYBREAK,
 };
 
 enum tunable_type_id {
diff --git a/net/core/ethtool.c b/net/core/ethtool.c
index 27e61b8..1600aa2 100644
--- a/net/core/ethtool.c
+++ b/net/core/ethtool.c
@@ -1625,6 +1625,7 @@ static int ethtool_tunable_valid(const struct ethtool_tunable *tuna)
 {
 	switch (tuna->id) {
 	case ETHTOOL_RX_COPYBREAK:
+	case ETHTOOL_TX_COPYBREAK:
 		if (tuna->len != sizeof(u32) ||
 		    tuna->type_id != ETHTOOL_TUNABLE_U32)
 			return -EINVAL;
-- 
1.8.3.4

[PATCH net-next 14/14] net/mlx4_en: Use the new tx_copybreak to set inline threshold

[Amir Vadai]

From: Eric Dumazet <edumazet@google.com>

Instead of setting inline threshold using module parameter only on
driver load, use set_tunable() to set it dynamically.
No need to store the threshold per ring, using instead the netdev global
priv->prof->inline_thold
Initial value still is set using the module parameter, therefore
backward compatability is kept.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_ethtool.c | 44 +++++++++++++++++++++++++
 drivers/net/ethernet/mellanox/mlx4/en_tx.c      |  1 -
 drivers/net/ethernet/mellanox/mlx4/mlx4_en.h    |  1 -
 3 files changed, 44 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
index 42c9f8b..ae83da9 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
@@ -1267,6 +1267,48 @@ static u32 mlx4_en_get_priv_flags(struct net_device *dev)
 	return priv->pflags;
 }
 
+static int mlx4_en_get_tunable(struct net_device *dev,
+			       const struct ethtool_tunable *tuna,
+			       void *data)
+{
+	const struct mlx4_en_priv *priv = netdev_priv(dev);
+	int ret = 0;
+
+	switch (tuna->id) {
+	case ETHTOOL_TX_COPYBREAK:
+		*(u32 *)data = priv->prof->inline_thold;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
+static int mlx4_en_set_tunable(struct net_device *dev,
+			       const struct ethtool_tunable *tuna,
+			       const void *data)
+{
+	struct mlx4_en_priv *priv = netdev_priv(dev);
+	int val, ret = 0;
+
+	switch (tuna->id) {
+	case ETHTOOL_TX_COPYBREAK:
+		val = *(u32 *)data;
+		if (val < MIN_PKT_LEN || val > MAX_INLINE)
+			ret = -EINVAL;
+		else
+			priv->prof->inline_thold = val;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	return ret;
+}
+
 
 const struct ethtool_ops mlx4_en_ethtool_ops = {
 	.get_drvinfo = mlx4_en_get_drvinfo,
@@ -1297,6 +1339,8 @@ const struct ethtool_ops mlx4_en_ethtool_ops = {
 	.get_ts_info = mlx4_en_get_ts_info,
 	.set_priv_flags = mlx4_en_set_priv_flags,
 	.get_priv_flags = mlx4_en_get_priv_flags,
+	.get_tunable		= mlx4_en_get_tunable,
+	.set_tunable		= mlx4_en_set_tunable,
 };
 
 
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index f0080c5..92a7cf4 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -66,7 +66,6 @@ int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
 	ring->size = size;
 	ring->size_mask = size - 1;
 	ring->stride = stride;
-	ring->inline_thold = priv->prof->inline_thold;
 
 	tmp = size * sizeof(struct mlx4_en_tx_info);
 	ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index a904030..8fef658 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -295,7 +295,6 @@ struct mlx4_en_tx_ring {
 	bool			bf_alloced;
 	struct netdev_queue	*tx_queue;
 	int			hwtstamp_tx_type;
-	int			inline_thold;
 } ____cacheline_aligned_in_smp;
 
 struct mlx4_en_rx_desc {
-- 
1.8.3.4

Re: [PATCH net-next 00/14] net/mlx4_en: Optimizations to TX flow

[Amir Vadai]

On 10/5/2014 12:35 PM, Amir Vadai wrote:
> I am sending this patchset now since the merge window is near and don't want to
> miss it.

A small clarification, since I didn't explain myself precisely enough:
This patchset is not a WIP and ready for merge. The open issues
aren't directly related to the series and will be addressed in
incremental manner.

Amir

Re: [PATCH net-next 14/14] net/mlx4_en: Use the new tx_copybreak to set inline threshold

[Sergei Shtylyov]

Hello.

On 10/5/2014 1:35 PM, Amir Vadai wrote:

> From: Eric Dumazet <edumazet@google.com>

> Instead of setting inline threshold using module parameter only on
> driver load, use set_tunable() to set it dynamically.
> No need to store the threshold per ring, using instead the netdev global
> priv->prof->inline_thold
> Initial value still is set using the module parameter, therefore
> backward compatability is kept.

> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Signed-off-by: Amir Vadai <amirv@mellanox.com>
> ---
>   drivers/net/ethernet/mellanox/mlx4/en_ethtool.c | 44 +++++++++++++++++++++++++
>   drivers/net/ethernet/mellanox/mlx4/en_tx.c      |  1 -
>   drivers/net/ethernet/mellanox/mlx4/mlx4_en.h    |  1 -
>   3 files changed, 44 insertions(+), 2 deletions(-)

> diff --git a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
> index 42c9f8b..ae83da9 100644
> --- a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
> +++ b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
[...]
> @@ -1297,6 +1339,8 @@ const struct ethtool_ops mlx4_en_ethtool_ops = {
>   	.get_ts_info = mlx4_en_get_ts_info,
>   	.set_priv_flags = mlx4_en_set_priv_flags,
>   	.get_priv_flags = mlx4_en_get_priv_flags,
> +	.get_tunable		= mlx4_en_get_tunable,
> +	.set_tunable		= mlx4_en_set_tunable,

    If the above initializers aren't aligned with tabs, why should these two be?

WBR, Sergei

Re: [PATCH net-next 00/14] net/mlx4_en: Optimizations to TX flow

[Eric Dumazet]

On Sun, 2014-10-05 at 14:45 +0300, Amir Vadai wrote:
> On 10/5/2014 12:35 PM, Amir Vadai wrote:
> > I am sending this patchset now since the merge window is near and don't want to
> > miss it.
> 
> A small clarification, since I didn't explain myself precisely enough:
> This patchset is not a WIP and ready for merge. The open issues
> aren't directly related to the series and will be addressed in
> incremental manner.


+2

Thanks Amir for this hard work.

Note that now the validate calls (GSO segmentation) is performed outside
of qdisc or device lock ( 55a93b3ea780 qdisc: validate skb without
holding lock) , its very easy to add a dev->gso_min_segs

I already have a patch for this, I am sending it right now.

Thanks !

[PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

From: Eric Dumazet <edumazet@google.com>

Some TSO engines might have a too heavy setup cost, that impacts
performance on hosts sending small bursts (2 MSS per packet).

This patch adds a device gso_min_segs, allowing drivers to set
a minimum segment size for TSO packets, according to the NIC
performance.

Tested on a mlx4 NIC, this allows to get a ~110% increase of
throughput when sending 2 MSS per packet.

Signed-off-by: Eric Dumazet <edumazet@google.com>
---
mlx4 patch will be sent later, its a one liner.

 include/linux/netdevice.h |    4 +++-
 net/core/dev.c            |    9 ++++++---
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 22d54b9b700d..2df86f50261c 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1416,6 +1416,8 @@ enum netdev_priv_flags {
  *	@gso_max_size:	Maximum size of generic segmentation offload
  *	@gso_max_segs:	Maximum number of segments that can be passed to the
  *			NIC for GSO
+ *	@gso_min_segs:	Minimum number of segments that can be passed to the
+ *			NIC for GSO
  *
  *	@dcbnl_ops:	Data Center Bridging netlink ops
  *	@num_tc:	Number of traffic classes in the net device
@@ -1666,7 +1668,7 @@ struct net_device {
 	unsigned int		gso_max_size;
 #define GSO_MAX_SEGS		65535
 	u16			gso_max_segs;
-
+	u16			gso_min_segs;
 #ifdef CONFIG_DCB
 	const struct dcbnl_rtnl_ops *dcbnl_ops;
 #endif
diff --git a/net/core/dev.c b/net/core/dev.c
index 1a90530f83ff..16e8ebbd3316 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -2567,10 +2567,12 @@ static netdev_features_t harmonize_features(struct sk_buff *skb,
 
 netdev_features_t netif_skb_features(struct sk_buff *skb)
 {
+	const struct net_device *dev = skb->dev;
+	netdev_features_t features = dev->features;
+	u16 gso_segs = skb_shinfo(skb)->gso_segs;
 	__be16 protocol = skb->protocol;
-	netdev_features_t features = skb->dev->features;
 
-	if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
+	if (gso_segs > dev->gso_max_segs || gso_segs < dev->gso_min_segs)
 		features &= ~NETIF_F_GSO_MASK;
 
 	if (protocol == htons(ETH_P_8021Q) || protocol == htons(ETH_P_8021AD)) {
@@ -2581,7 +2583,7 @@ netdev_features_t netif_skb_features(struct sk_buff *skb)
 	}
 
 	features = netdev_intersect_features(features,
-					     skb->dev->vlan_features |
+					     dev->vlan_features |
 					     NETIF_F_HW_VLAN_CTAG_TX |
 					     NETIF_F_HW_VLAN_STAG_TX);
 
@@ -6658,6 +6660,7 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
 
 	dev->gso_max_size = GSO_MAX_SIZE;
 	dev->gso_max_segs = GSO_MAX_SEGS;
+	dev->gso_min_segs = 0;
 
 	INIT_LIST_HEAD(&dev->napi_list);
 	INIT_LIST_HEAD(&dev->unreg_list);

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Tom Herbert]

On Sun, Oct 5, 2014 at 10:11 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> From: Eric Dumazet <edumazet@google.com>
>
> Some TSO engines might have a too heavy setup cost, that impacts
> performance on hosts sending small bursts (2 MSS per packet).
>
> This patch adds a device gso_min_segs, allowing drivers to set
> a minimum segment size for TSO packets, according to the NIC
> performance.
>
Eric, this seems like another device specific limitation in TSO we are
exposing to the stack. Can't we just put things like this in
ndo_gso_check so driver can apply whatever criteria it wants in
deciding if it worth it or even to rather do TSO?

> Tested on a mlx4 NIC, this allows to get a ~110% increase of
> throughput when sending 2 MSS per packet.
>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> ---
> mlx4 patch will be sent later, its a one liner.
>
>  include/linux/netdevice.h |    4 +++-
>  net/core/dev.c            |    9 ++++++---
>  2 files changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index 22d54b9b700d..2df86f50261c 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h
> @@ -1416,6 +1416,8 @@ enum netdev_priv_flags {
>   *     @gso_max_size:  Maximum size of generic segmentation offload
>   *     @gso_max_segs:  Maximum number of segments that can be passed to the
>   *                     NIC for GSO
> + *     @gso_min_segs:  Minimum number of segments that can be passed to the
> + *                     NIC for GSO
>   *
>   *     @dcbnl_ops:     Data Center Bridging netlink ops
>   *     @num_tc:        Number of traffic classes in the net device
> @@ -1666,7 +1668,7 @@ struct net_device {
>         unsigned int            gso_max_size;
>  #define GSO_MAX_SEGS           65535
>         u16                     gso_max_segs;
> -
> +       u16                     gso_min_segs;
>  #ifdef CONFIG_DCB
>         const struct dcbnl_rtnl_ops *dcbnl_ops;
>  #endif
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1a90530f83ff..16e8ebbd3316 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2567,10 +2567,12 @@ static netdev_features_t harmonize_features(struct sk_buff *skb,
>
>  netdev_features_t netif_skb_features(struct sk_buff *skb)
>  {
> +       const struct net_device *dev = skb->dev;
> +       netdev_features_t features = dev->features;
> +       u16 gso_segs = skb_shinfo(skb)->gso_segs;
>         __be16 protocol = skb->protocol;
> -       netdev_features_t features = skb->dev->features;
>
> -       if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
> +       if (gso_segs > dev->gso_max_segs || gso_segs < dev->gso_min_segs)
>                 features &= ~NETIF_F_GSO_MASK;
>
>         if (protocol == htons(ETH_P_8021Q) || protocol == htons(ETH_P_8021AD)) {
> @@ -2581,7 +2583,7 @@ netdev_features_t netif_skb_features(struct sk_buff *skb)
>         }
>
>         features = netdev_intersect_features(features,
> -                                            skb->dev->vlan_features |
> +                                            dev->vlan_features |
>                                              NETIF_F_HW_VLAN_CTAG_TX |
>                                              NETIF_F_HW_VLAN_STAG_TX);
>
> @@ -6658,6 +6660,7 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
>
>         dev->gso_max_size = GSO_MAX_SIZE;
>         dev->gso_max_segs = GSO_MAX_SEGS;
> +       dev->gso_min_segs = 0;
>
>         INIT_LIST_HEAD(&dev->napi_list);
>         INIT_LIST_HEAD(&dev->unreg_list);
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Sun, 2014-10-05 at 11:45 -0700, Tom Herbert wrote:
> On Sun, Oct 5, 2014 at 10:11 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > From: Eric Dumazet <edumazet@google.com>
> >
> > Some TSO engines might have a too heavy setup cost, that impacts
> > performance on hosts sending small bursts (2 MSS per packet).
> >
> > This patch adds a device gso_min_segs, allowing drivers to set
> > a minimum segment size for TSO packets, according to the NIC
> > performance.
> >
> Eric, this seems like another device specific limitation in TSO we are
> exposing to the stack. Can't we just put things like this in
> ndo_gso_check so driver can apply whatever criteria it wants in
> deciding if it worth it or even to rather do TSO?

I don't think it really matters where the check is done.

If you want to move the check away from netif_skb_features(), its fine
with me, but please note you did not suggest this for current check
against gso_max_segs.

I based my patch on current net-next, where ndo_gso_check() does not
exist.

BTW, I wonder now why we added gso_max_segs support to bonding driver.
(commit 0e376bd0b791ac6ac6bdb051492df0769c840848)

It makes little sense, gso should be done at the last possible stage.

Re: [PATCH net-next 00/14] net/mlx4_en: Optimizations to TX flow

[David Miller]

From: Amir Vadai <amirv@mellanox.com>
Date: Sun,  5 Oct 2014 12:35:08 +0300

> This patchset contains optimizations to TX flow in mlx4_en driver. It also introduce
> setting/getting tx copybreak, to enable controlling inline threshold dynamically.
> 
> TX flow optimizations was authored and posted to the mailing list by Eric
> Dumazet [1] as a single patch. I splitted this patch to smaller patches,
> Reviewed it and tested.
> Changed from original patch:
> - s/iowrite32be/iowrite32/, since ring->doorbell_qpn is stored as be32
> 
> The tx copybreak patch was also suggested by Eric Dumazet, and was edited and
> reviewed by me. User space patch will be sent after kernel code is ready.
> 
> I am sending this patchset now since the merge window is near and don't want to
> miss it.
> 
> More work need to do:
> - Disable BF when xmit_more is in use
> - Make TSO use xmit_more too. Maybe by splitting small TSO packets in the
>   driver itself, to avoid extra cpu/memory costs of GSO before the driver
> - Fix mlx4_en_xmit buggy handling of queue full in the middle of a burst
>   partially posted to send queue using xmit_more
> 
> Eric, I edited the patches to have you as the Author and the first
> signed-off-by. I hope it is ok with you (I wasn't sure if it is ok to sign by
> you), anyway all the credit to those changes should go to you.
> 
> Patchset was tested and applied over commit 1e203c1 "(net: sched:
> suspicious RCU usage in qdisc_watchdog")
> 
> [1] - https://patchwork.ozlabs.org/patch/394256/

Looks great, nice work everyone.

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Amir Vadai]

On 10/5/2014 8:11 PM, Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
> 
> Some TSO engines might have a too heavy setup cost, that impacts
> performance on hosts sending small bursts (2 MSS per packet).
> 
> This patch adds a device gso_min_segs, allowing drivers to set
> a minimum segment size for TSO packets, according to the NIC
> performance.
> 
> Tested on a mlx4 NIC, this allows to get a ~110% increase of
> throughput when sending 2 MSS per packet.
> 

Amazing!

Shouldn't there be a netif_set_gso_min_size() too?

> Signed-off-by: Eric Dumazet <edumazet@google.com>
> ---
> mlx4 patch will be sent later, its a one liner.
> 
>  include/linux/netdevice.h |    4 +++-
>  net/core/dev.c            |    9 ++++++---
>  2 files changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index 22d54b9b700d..2df86f50261c 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h
> @@ -1416,6 +1416,8 @@ enum netdev_priv_flags {
>   *	@gso_max_size:	Maximum size of generic segmentation offload
>   *	@gso_max_segs:	Maximum number of segments that can be passed to the
>   *			NIC for GSO
> + *	@gso_min_segs:	Minimum number of segments that can be passed to the
> + *			NIC for GSO
>   *
>   *	@dcbnl_ops:	Data Center Bridging netlink ops
>   *	@num_tc:	Number of traffic classes in the net device
> @@ -1666,7 +1668,7 @@ struct net_device {
>  	unsigned int		gso_max_size;
>  #define GSO_MAX_SEGS		65535
>  	u16			gso_max_segs;
> -
> +	u16			gso_min_segs;
>  #ifdef CONFIG_DCB
>  	const struct dcbnl_rtnl_ops *dcbnl_ops;
>  #endif
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1a90530f83ff..16e8ebbd3316 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2567,10 +2567,12 @@ static netdev_features_t harmonize_features(struct sk_buff *skb,
>  
>  netdev_features_t netif_skb_features(struct sk_buff *skb)
>  {
> +	const struct net_device *dev = skb->dev;
> +	netdev_features_t features = dev->features;
> +	u16 gso_segs = skb_shinfo(skb)->gso_segs;
>  	__be16 protocol = skb->protocol;
> -	netdev_features_t features = skb->dev->features;
>  
> -	if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
> +	if (gso_segs > dev->gso_max_segs || gso_segs < dev->gso_min_segs)
>  		features &= ~NETIF_F_GSO_MASK;
>  
>  	if (protocol == htons(ETH_P_8021Q) || protocol == htons(ETH_P_8021AD)) {
> @@ -2581,7 +2583,7 @@ netdev_features_t netif_skb_features(struct sk_buff *skb)
>  	}
>  
>  	features = netdev_intersect_features(features,
> -					     skb->dev->vlan_features |
> +					     dev->vlan_features |
>  					     NETIF_F_HW_VLAN_CTAG_TX |
>  					     NETIF_F_HW_VLAN_STAG_TX);
>  
> @@ -6658,6 +6660,7 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
>  
>  	dev->gso_max_size = GSO_MAX_SIZE;
>  	dev->gso_max_segs = GSO_MAX_SEGS;
> +	dev->gso_min_segs = 0;
>  
>  	INIT_LIST_HEAD(&dev->napi_list);
>  	INIT_LIST_HEAD(&dev->unreg_list);
> 
> 

RE: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[David Laight]

From: Eric Dumazet <edumazet@google.com>
> Some TSO engines might have a too heavy setup cost, that impacts
> performance on hosts sending small bursts (2 MSS per packet).
> 
> This patch adds a device gso_min_segs, allowing drivers to set
> a minimum segment size for TSO packets, according to the NIC
> performance.
> 
> Tested on a mlx4 NIC, this allows to get a ~110% increase of
> throughput when sending 2 MSS per packet.

Doesn't this all depend on what you need to optimise for.
I can think of three^Wseveral main cases:
1) minimising cpu use while saturating the local network.
2) minimising latency for single packets.
3) maximising throughput for a single connection.
4) minimising cpu use when handling a large number of connections.
plus all the variations in packet size.

I'm not sure what you are trading for what here.

Maybe gso = tx_bursting is almost always faster on some hardware?
(Especially if an idle mac engine is 'kicked' for the first packet
of a burst.)

	David

RE: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Mon, 2014-10-06 at 10:20 +0000, David Laight wrote:
> From: Eric Dumazet <edumazet@google.com>
> > Some TSO engines might have a too heavy setup cost, that impacts
> > performance on hosts sending small bursts (2 MSS per packet).
> > 
> > This patch adds a device gso_min_segs, allowing drivers to set
> > a minimum segment size for TSO packets, according to the NIC
> > performance.
> > 
> > Tested on a mlx4 NIC, this allows to get a ~110% increase of
> > throughput when sending 2 MSS per packet.
> 
> Doesn't this all depend on what you need to optimise for.
> I can think of three^Wseveral main cases:
> 1) minimising cpu use while saturating the local network.
> 2) minimising latency for single packets.
> 3) maximising throughput for a single connection.
> 4) minimising cpu use when handling a large number of connections.
> plus all the variations in packet size.
> 
> I'm not sure what you are trading for what here.
> 

I am not sure you really understood.

What's the point having a 40GB NIC and not being able to reach 18 Gb on
it, even if you are willing to spend all the cpu cycles you want ?


> Maybe gso = tx_bursting is almost always faster on some hardware?
> (Especially if an idle mac engine is 'kicked' for the first packet
> of a burst.)

This has nothing to do with xmit_more.

I start 1200 flows rate limited to 3 MBytes each.
(1000 netperf -t TCP_STREAM, nothing fancy here)

Theoretical total of 3.6 GBytes per second.

Without patch :
# sar -n DEV 5 5 | grep eth0
05:07:56 AM      eth0 555621.60 1111203.20  35813.03 1642923.46      0.00      0.00      0.60
05:08:01 AM      eth0 555591.00 1111173.80  35810.47 1642877.52      0.00      0.00      0.40
05:08:06 AM      eth0 555586.20 1111162.60  35810.06 1642861.03      0.00      0.00      0.60
05:08:11 AM      eth0 555624.40 1111235.40  35812.75 1642974.19      0.00      0.00      0.60
05:08:16 AM      eth0 555639.60 1111266.80  35813.21 1643017.83      0.00      0.00      0.60
Average:         eth0 555612.56 1111208.36  35811.90 1642930.81      0.00      0.00      0.56

With patch :

# sar -n DEV 5 5 | grep eth0
05:07:04 AM      eth0 1179478.80 2358940.40  76022.47 3487725.22      0.00      0.00      0.60
05:07:09 AM      eth0 1178913.60 2357807.40  75986.60 3486044.00      0.00      0.00      0.40
05:07:14 AM      eth0 1178957.40 2357897.60  75988.98 3486177.50      0.00      0.00      0.60
05:07:19 AM      eth0 1177556.00 2355064.60  75899.37 3481993.37      0.00      0.00      0.60
05:07:24 AM      eth0 1180321.20 2360625.20  76077.15 3490209.58      0.00      0.00      0.40
Average:         eth0 1179045.40 2358067.04  75994.92 3486429.94      0.00      0.00      0.52

Ask yourself which one we prefer.

About cpu costs, we hardly see anything caused by GSO,
now we optimized __copy_skb_header()

     6.36%        swapper  [kernel.kallsyms]   [k] _raw_spin_lock                            
     5.24%        netperf  [kernel.kallsyms]   [k] copy_user_enhanced_fast_string            
     5.03%        swapper  [kernel.kallsyms]   [k] poll_idle                                 
     3.73%        swapper  [kernel.kallsyms]   [k] tcp_ack                                   
     2.73%        swapper  [kernel.kallsyms]   [k] memcpy                                    
     2.49%        swapper  [kernel.kallsyms]   [k] __skb_clone                               
     2.41%        swapper  [kernel.kallsyms]   [k] skb_release_data                          
     2.33%        swapper  [kernel.kallsyms]   [k] intel_idle                                
     2.23%        swapper  [kernel.kallsyms]   [k] tcp_init_tso_segs                         
     1.99%        swapper  [kernel.kallsyms]   [k] fq_dequeue                                
     1.94%        netperf  [kernel.kallsyms]   [k] tcp_sendmsg                               
     1.82%        swapper  [kernel.kallsyms]   [k] tcp_write_xmit                            
     1.28%        swapper  [kernel.kallsyms]   [k] __netif_receive_skb_core                  
     1.23%        swapper  [kernel.kallsyms]   [k] __copy_skb_header                         
     1.14%        swapper  [kernel.kallsyms]   [k] kfree                                     
     1.10%        swapper  [kernel.kallsyms]   [k] kmem_cache_free                           
     1.06%        swapper  [kernel.kallsyms]   [k] mlx4_en_xmit                              
     1.05%        swapper  [kernel.kallsyms]   [k] tcp_wfree                                 
     1.02%        swapper  [kernel.kallsyms]   [k] inet_gso_segment                          
     1.01%        swapper  [kernel.kallsyms]   [k] put_compound_page                         
     0.98%        swapper  [kernel.kallsyms]   [k] _raw_spin_lock_irqsave                    
     0.96%        netperf  [kernel.kallsyms]   [k] __alloc_skb                               
     0.92%        netperf  [kernel.kallsyms]   [k] _raw_spin_lock                            
     0.89%        swapper  [kernel.kallsyms]   [k] skb_segment                               
     0.88%        swapper  [kernel.kallsyms]   [k] tcp_transmit_skb                          
     0.82%        swapper  [kernel.kallsyms]   [k] ip_queue_xmit                             
     0.81%        swapper  [kernel.kallsyms]   [k] __inet_lookup_established                 
     0.76%        swapper  [kernel.kallsyms]   [k] __kfree_skb                               
     0.73%        swapper  [kernel.kallsyms]   [k] ipv4_dst_check                            
     0.66%        netperf  [kernel.kallsyms]   [k] tcp_ack                                   
     0.60%        swapper  [kernel.kallsyms]   [k] __alloc_skb                               
     0.56%        swapper  [kernel.kallsyms]   [k] ipt_do_table                              

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Mon, 2014-10-06 at 09:41 +0300, Amir Vadai wrote:
> On 10/5/2014 8:11 PM, Eric Dumazet wrote:
> > From: Eric Dumazet <edumazet@google.com>
> > 
> > Some TSO engines might have a too heavy setup cost, that impacts
> > performance on hosts sending small bursts (2 MSS per packet).
> > 
> > This patch adds a device gso_min_segs, allowing drivers to set
> > a minimum segment size for TSO packets, according to the NIC
> > performance.
> > 
> > Tested on a mlx4 NIC, this allows to get a ~110% increase of
> > throughput when sending 2 MSS per packet.
> > 
> 
> Amazing!
> 
> Shouldn't there be a netif_set_gso_min_size() too?

Good point, I'll add it in v2, thanks !

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Mon, 2014-10-06 at 05:17 -0700, Eric Dumazet wrote:
> On Mon, 2014-10-06 at 09:41 +0300, Amir Vadai wrote:
> > On 10/5/2014 8:11 PM, Eric Dumazet wrote:
> > > From: Eric Dumazet <edumazet@google.com>
> > > 
> > > Some TSO engines might have a too heavy setup cost, that impacts
> > > performance on hosts sending small bursts (2 MSS per packet).
> > > 
> > > This patch adds a device gso_min_segs, allowing drivers to set
> > > a minimum segment size for TSO packets, according to the NIC
> > > performance.
> > > 
> > > Tested on a mlx4 NIC, this allows to get a ~110% increase of
> > > throughput when sending 2 MSS per packet.
> > > 
> > 
> > Amazing!
> > 
> > Shouldn't there be a netif_set_gso_min_size() too?
> 
> Good point, I'll add it in v2, thanks !

Well, no... I thought you were speaking of an hypothetical
netif_set_gso_{max|min}_segs()

But we don't have it. : Drivers simply change dev->gso_max_segs

drivers/net/ethernet/freescale/fec_main.c:3058:         ndev->gso_max_segs = FEC_MAX_TSO_SEGS;
drivers/net/ethernet/marvell/mv643xx_eth.c:3120:        dev->gso_max_segs = MV643XX_MAX_TSO_SEGS;
drivers/net/ethernet/marvell/mvneta.c:3057:     dev->gso_max_segs = MVNETA_MAX_TSO_SEGS;
drivers/net/ethernet/sfc/efx.c:2300:    net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;

So my mlx4 patch was simply doing

	dev->gso_min_segs = MLX4_MIN_TSO_SEGS;

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Sun, 05 Oct 2014 10:11:27 -0700

> From: Eric Dumazet <edumazet@google.com>
> 
> Some TSO engines might have a too heavy setup cost, that impacts
> performance on hosts sending small bursts (2 MSS per packet).
> 
> This patch adds a device gso_min_segs, allowing drivers to set
> a minimum segment size for TSO packets, according to the NIC
> performance.
> 
> Tested on a mlx4 NIC, this allows to get a ~110% increase of
> throughput when sending 2 MSS per packet.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>

So exactly what value are you using for mlx4?

Because I wonder if we should just generically forfeit TSO unless
we have > 2 segments, for example.

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Mon, 2014-10-06 at 17:21 -0400, David Miller wrote:

> So exactly what value are you using for mlx4?
> 

It seems that on ConnectX-3 family, TSO packets of 2 or 3 MSS are not
worth using TSO engine. The cutoff point seems to be 4 (same throughput)

So I was planning to use gso_min_segs = 4 only for them.

> Because I wonder if we should just generically forfeit TSO unless
> we have > 2 segments, for example.

When I tested on bnx2x, this was not a gain.

bnx2x is faster sending TSO packets, even if they have 2 MSS.

I'll try the experiment on I40E Intel cards.

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Mon, 06 Oct 2014 14:42:58 -0700

> On Mon, 2014-10-06 at 17:21 -0400, David Miller wrote:
> 
>> So exactly what value are you using for mlx4?
>> 
> 
> It seems that on ConnectX-3 family, TSO packets of 2 or 3 MSS are not
> worth using TSO engine. The cutoff point seems to be 4 (same throughput)
> 
> So I was planning to use gso_min_segs = 4 only for them.
> 
>> Because I wonder if we should just generically forfeit TSO unless
>> we have > 2 segments, for example.
> 
> When I tested on bnx2x, this was not a gain.
> 
> bnx2x is faster sending TSO packets, even if they have 2 MSS.
> 
> I'll try the experiment on I40E Intel cards.

Ok I'm sold on your patch then if two major chipsets already benefit
from differing values.

I'll apply this, thanks Eric.

Re: [PATCH net-next] net: introduce netdevice gso_min_segs attribute

[Eric Dumazet]

On Mon, 2014-10-06 at 17:54 -0400, David Miller wrote:

> Ok I'm sold on your patch then if two major chipsets already benefit
> from differing values.
> 
> I'll apply this, thanks Eric.

I did the experiment on I40e, and no difference on this NIC.

(Note it stills lacks xmit_more support atm)

Thanks David