bnx2x: uses build_skb() in receive path

bnx2x uses following formula to compute its rx_buf_sz :

dev->mtu + 2*L1_CACHE_BYTES + 14 + 8 + 8 + 2

Then core network adds NET_SKB_PAD and SKB_DATA_ALIGN(sizeof(struct
skb_shared_info))

Final allocated size for skb head on x86_64 (L1_CACHE_BYTES = 64,
MTU=1500) : 2112 bytes : SLUB/SLAB round this to 4096 bytes.

Since skb truesize is then bigger than SK_MEM_QUANTUM, we have lot of
false sharing because of mem_reclaim in UDP stack.

One possible way to half truesize is to reduce the need by 64 bytes
(2112 -> 2048 bytes)

Instead of allocating a full cache line at the end of packet for
alignment, we can use the fact that skb_shared_info sits at the end of
skb->head, and we can use this room, if we convert bnx2x to new
build_skb() infrastructure.

skb_shared_info will be initialized after hardware finished its
transfert, so we can eventually overwrite the final padding.

Using build_skb() also reduces cache line misses in the driver, since we
use cache hot skb instead of cold ones. Number of in-flight sk_buff
structures is lower, they are recycled while still hot.

Performance results :

(820.000 pps on a rx UDP monothread benchmark, instead of 720.000 pps)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Acked-by: Eilon Greenstein <eilong@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnx2x/bnx2x.h

@@ -293,8 +293,13 @@ enum {
 #define FCOE_TXQ_IDX(bp)	(MAX_ETH_TXQ_IDX(bp))
 
 /* fast path */
+/*
+ * This driver uses new build_skb() API :
+ * RX ring buffer contains pointer to kmalloc() data only,
+ * skb are built only after Hardware filled the frame.
+ */
 struct sw_rx_bd {
-	struct sk_buff	*skb;
+	u8		*data;
 	DEFINE_DMA_UNMAP_ADDR(mapping);
 };
 
@@ -424,8 +429,8 @@ union host_hc_status_block {
 
 struct bnx2x_agg_info {
 	/*
-	 * First aggregation buffer is an skb, the following - are pages.
-	 * We will preallocate the skbs for each aggregation when
+	 * First aggregation buffer is a data buffer, the following - are pages.
+	 * We will preallocate the data buffer for each aggregation when
 	 * we open the interface and will replace the BD at the consumer
 	 * with this one when we receive the TPA_START CQE in order to
 	 * keep the Rx BD ring consistent.
@@ -439,6 +444,7 @@ struct bnx2x_agg_info {
 	u16			parsing_flags;
 	u16			vlan_tag;
 	u16			len_on_bd;
+	u32			rxhash;
 };
 
 #define Q_STATS_OFFSET32(stat_name) \
@@ -1187,10 +1193,20 @@ struct bnx2x {
 #define ETH_MAX_JUMBO_PACKET_SIZE	9600
 
 	/* Max supported alignment is 256 (8 shift) */
-#define BNX2X_RX_ALIGN_SHIFT		((L1_CACHE_SHIFT < 8) ? \
-					 L1_CACHE_SHIFT : 8)
-	/* FW use 2 Cache lines Alignment for start packet and size  */
-#define BNX2X_FW_RX_ALIGN		(2 << BNX2X_RX_ALIGN_SHIFT)
+#define BNX2X_RX_ALIGN_SHIFT		min(8, L1_CACHE_SHIFT)
+
+	/* FW uses 2 Cache lines Alignment for start packet and size
+	 *
+	 * We assume skb_build() uses sizeof(struct skb_shared_info) bytes
+	 * at the end of skb->data, to avoid wasting a full cache line.
+	 * This reduces memory use (skb->truesize).
+	 */
+#define BNX2X_FW_RX_ALIGN_START	(1UL << BNX2X_RX_ALIGN_SHIFT)
+
+#define BNX2X_FW_RX_ALIGN_END					\
+	max(1UL << BNX2X_RX_ALIGN_SHIFT, 			\
+	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
+
 #define BNX2X_PXP_DRAM_ALIGN		(BNX2X_RX_ALIGN_SHIFT - 5)
 
 	struct host_sp_status_block *def_status_blk;

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.h

@@ -910,26 +910,27 @@ static inline int bnx2x_alloc_rx_sge(struct bnx2x *bp,
 	return 0;
 }
 
-static inline int bnx2x_alloc_rx_skb(struct bnx2x *bp,
+static inline int bnx2x_alloc_rx_data(struct bnx2x *bp,
 				      struct bnx2x_fastpath *fp, u16 index)
 {
-	struct sk_buff *skb;
+	u8 *data;
 	struct sw_rx_bd *rx_buf = &fp->rx_buf_ring[index];
 	struct eth_rx_bd *rx_bd = &fp->rx_desc_ring[index];
 	dma_addr_t mapping;
 
-	skb = netdev_alloc_skb(bp->dev, fp->rx_buf_size);
-	if (unlikely(skb == NULL))
+	data = kmalloc(fp->rx_buf_size + NET_SKB_PAD, GFP_ATOMIC);
+	if (unlikely(data == NULL))
 		return -ENOMEM;
 
-	mapping = dma_map_single(&bp->pdev->dev, skb->data, fp->rx_buf_size,
+	mapping = dma_map_single(&bp->pdev->dev, data + NET_SKB_PAD,
+				 fp->rx_buf_size,
 				 DMA_FROM_DEVICE);
 	if (unlikely(dma_mapping_error(&bp->pdev->dev, mapping))) {
-		dev_kfree_skb_any(skb);
+		kfree(data);
 		return -ENOMEM;
 	}
 
-	rx_buf->skb = skb;
+	rx_buf->data = data;
 	dma_unmap_addr_set(rx_buf, mapping, mapping);
 
 	rx_bd->addr_hi = cpu_to_le32(U64_HI(mapping));
@@ -938,12 +939,12 @@ static inline int bnx2x_alloc_rx_skb(struct bnx2x *bp,
 	return 0;
 }
 
-/* note that we are not allocating a new skb,
+/* note that we are not allocating a new buffer,
  * we are just moving one from cons to prod
  * we are not creating a new mapping,
  * so there is no need to check for dma_mapping_error().
  */
-static inline void bnx2x_reuse_rx_skb(struct bnx2x_fastpath *fp,
+static inline void bnx2x_reuse_rx_data(struct bnx2x_fastpath *fp,
 				      u16 cons, u16 prod)
 {
 	struct sw_rx_bd *cons_rx_buf = &fp->rx_buf_ring[cons];
@@ -953,7 +954,7 @@ static inline void bnx2x_reuse_rx_skb(struct bnx2x_fastpath *fp,
 
 	dma_unmap_addr_set(prod_rx_buf, mapping,
 			   dma_unmap_addr(cons_rx_buf, mapping));
-	prod_rx_buf->skb = cons_rx_buf->skb;
+	prod_rx_buf->data = cons_rx_buf->data;
 	*prod_bd = *cons_bd;
 }
 
@@ -1029,9 +1030,9 @@ static inline void bnx2x_free_tpa_pool(struct bnx2x *bp,
 	for (i = 0; i < last; i++) {
 		struct bnx2x_agg_info *tpa_info = &fp->tpa_info[i];
 		struct sw_rx_bd *first_buf = &tpa_info->first_buf;
-		struct sk_buff *skb = first_buf->skb;
+		u8 *data = first_buf->data;
 
-		if (skb == NULL) {
+		if (data == NULL) {
 			DP(NETIF_MSG_IFDOWN, "tpa bin %d empty on free\n", i);
 			continue;
 		}
@@ -1039,8 +1040,8 @@ static inline void bnx2x_free_tpa_pool(struct bnx2x *bp,
 			dma_unmap_single(&bp->pdev->dev,
 					 dma_unmap_addr(first_buf, mapping),
 					 fp->rx_buf_size, DMA_FROM_DEVICE);
-		dev_kfree_skb(skb);
-		first_buf->skb = NULL;
+		kfree(data);
+		first_buf->data = NULL;
 	}
 }
 
@@ -1148,7 +1149,7 @@ static inline int bnx2x_alloc_rx_bds(struct bnx2x_fastpath *fp,
 	 * fp->eth_q_stats.rx_skb_alloc_failed = 0
 	 */
 	for (i = 0; i < rx_ring_size; i++) {
-		if (bnx2x_alloc_rx_skb(bp, fp, ring_prod) < 0) {
+		if (bnx2x_alloc_rx_data(bp, fp, ring_prod) < 0) {
 			fp->eth_q_stats.rx_skb_alloc_failed++;
 			continue;
 		}

drivers/net/ethernet/broadcom/bnx2x/bnx2x_ethtool.c

@@ -1740,6 +1740,7 @@ static int bnx2x_run_loopback(struct bnx2x *bp, int loopback_mode)
 	struct sw_rx_bd *rx_buf;
 	u16 len;
 	int rc = -ENODEV;
+	u8 *data;
 
 	/* check the loopback mode */
 	switch (loopback_mode) {
@@ -1865,10 +1866,9 @@ static int bnx2x_run_loopback(struct bnx2x *bp, int loopback_mode)
 	dma_sync_single_for_cpu(&bp->pdev->dev,
 				   dma_unmap_addr(rx_buf, mapping),
 				   fp_rx->rx_buf_size, DMA_FROM_DEVICE);
-	skb = rx_buf->skb;
-	skb_reserve(skb, cqe->fast_path_cqe.placement_offset);
+	data = rx_buf->data + NET_SKB_PAD + cqe->fast_path_cqe.placement_offset;
 	for (i = ETH_HLEN; i < pkt_size; i++)
-		if (*(skb->data + i) != (unsigned char) (i & 0xff))
+		if (*(data + i) != (unsigned char) (i & 0xff))
 			goto test_loopback_rx_exit;
 
 	rc = 0;

drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c

@@ -2789,8 +2789,8 @@ static void bnx2x_pf_rx_q_prep(struct bnx2x *bp,
 	/* This should be a maximum number of data bytes that may be
 	 * placed on the BD (not including paddings).
 	 */
-	rxq_init->buf_sz = fp->rx_buf_size - BNX2X_FW_RX_ALIGN -
-		IP_HEADER_ALIGNMENT_PADDING;
+	rxq_init->buf_sz = fp->rx_buf_size - BNX2X_FW_RX_ALIGN_START -
+		BNX2X_FW_RX_ALIGN_END -	IP_HEADER_ALIGNMENT_PADDING;
 
 	rxq_init->cl_qzone_id = fp->cl_qzone_id;
 	rxq_init->tpa_agg_sz = tpa_agg_size;