qede: Add build_skb() support.

This patch makes use of build_skb() throughout in driver's receieve
data path [HW gro flow and non HW gro flow]. With this, driver can
build skb directly from the page segments which are already mapped
to the hardware instead of allocating new SKB via netdev_alloc_skb()
and memcpy the data which is quite costly.

This really improves performance (keeping same or slight gain in rx
throughput) in terms of CPU utilization which is significantly reduced
[almost half] in non HW gro flow where for every incoming MTU sized
packet driver had to allocate skb, memcpy headers etc. Additionally
in that flow, it also gets rid of bunch of additional overheads
[eth_get_headlen() etc.] to split headers and data in the skb.

Tested with:
system: 2 sockets, 4 cores per socket, hyperthreading, 2x4x2=16 cores
iperf [server]: iperf -s
iperf [client]: iperf -c <server_ip> -t 500 -i 10 -P 32

HW GRO off – w/o build_skb(), throughput: 36.8 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.59    0.00   32.93    0.00    0.00   43.07    0.00    0.00   23.42

HW GRO off - with build_skb(), throughput: 36.9 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.70    0.00   31.70    0.00    0.00   25.68    0.00    0.00   41.92

HW GRO on - w/o build_skb(), throughput: 36.9 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.86    0.00   24.14    0.00    0.00    6.59    0.00    0.00   68.41

HW GRO on - with build_skb(), throughput: 37.5 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.87    0.00   23.75    0.00    0.00    6.19    0.00    0.00   69.19
Signed-off-by: Ariel Elior <ariel.elior@cavium.com>
Signed-off-by: Manish Chopra <manish.chopra@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/qlogic/qede/qede.h

@@ -290,15 +290,12 @@ struct qede_agg_info {
 	 * aggregation.
 	 */
 	struct sw_rx_data buffer;
-	dma_addr_t buffer_mapping;
-
 	struct sk_buff *skb;
 
 	/* We need some structs from the start cookie until termination */
 	u16 vlan_tag;
-	u16 start_cqe_bd_len;
-	u8 start_cqe_placement_offset;
 
+	bool tpa_start_fail;
 	u8 state;
 	u8 frag_id;
 

drivers/net/ethernet/qlogic/qede/qede_ethtool.c

@@ -1508,7 +1508,8 @@ static int qede_selftest_receive_traffic(struct qede_dev *edev)
 		len =  le16_to_cpu(fp_cqe->len_on_first_bd);
 		data_ptr = (u8 *)(page_address(sw_rx_data->data) +
 				  fp_cqe->placement_offset +
-				  sw_rx_data->page_offset);
+				  sw_rx_data->page_offset +
+				  rxq->rx_headroom);
 		if (ether_addr_equal(data_ptr,  edev->ndev->dev_addr) &&
 		    ether_addr_equal(data_ptr + ETH_ALEN,
 				     edev->ndev->dev_addr)) {

drivers/net/ethernet/qlogic/qede/qede_fp.c

@@ -660,7 +660,8 @@ static int qede_fill_frag_skb(struct qede_dev *edev,
 
 	/* Add one frag and update the appropriate fields in the skb */
 	skb_fill_page_desc(skb, tpa_info->frag_id++,
-			   current_bd->data, current_bd->page_offset,
+			   current_bd->data,
+			   current_bd->page_offset + rxq->rx_headroom,
 			   len_on_bd);
 
 	if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
@@ -671,8 +672,7 @@ static int qede_fill_frag_skb(struct qede_dev *edev,
 		goto out;
 	}
 
-	qed_chain_consume(&rxq->rx_bd_ring);
-	rxq->sw_rx_cons++;
+	qede_rx_bd_ring_consume(rxq);
 
 	skb->data_len += len_on_bd;
 	skb->truesize += rxq->rx_buf_seg_size;
@@ -721,64 +721,129 @@ static u8 qede_check_tunn_csum(u16 flag)
 	return QEDE_CSUM_UNNECESSARY | tcsum;
 }
 
-static void qede_tpa_start(struct qede_dev *edev,
+static inline struct sk_buff *
+qede_build_skb(struct qede_rx_queue *rxq,
+	       struct sw_rx_data *bd, u16 len, u16 pad)
+{
+	struct sk_buff *skb;
+	void *buf;
+
+	buf = page_address(bd->data) + bd->page_offset;
+	skb = build_skb(buf, rxq->rx_buf_seg_size);
+
+	skb_reserve(skb, pad);
+	skb_put(skb, len);
+
+	return skb;
+}
+
+static struct sk_buff *
+qede_tpa_rx_build_skb(struct qede_dev *edev,
 		      struct qede_rx_queue *rxq,
-			   struct eth_fast_path_rx_tpa_start_cqe *cqe)
+		      struct sw_rx_data *bd, u16 len, u16 pad,
+		      bool alloc_skb)
 {
-	struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
-	struct eth_rx_bd *rx_bd_cons = qed_chain_consume(&rxq->rx_bd_ring);
-	struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
-	struct sw_rx_data *replace_buf = &tpa_info->buffer;
-	dma_addr_t mapping = tpa_info->buffer_mapping;
-	struct sw_rx_data *sw_rx_data_cons;
-	struct sw_rx_data *sw_rx_data_prod;
+	struct sk_buff *skb;
 
-	sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
-	sw_rx_data_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
+	skb = qede_build_skb(rxq, bd, len, pad);
+	bd->page_offset += rxq->rx_buf_seg_size;
 
-	/* Use pre-allocated replacement buffer - we can't release the agg.
-	 * start until its over and we don't want to risk allocation failing
-	 * here, so re-allocate when aggregation will be over.
-	 */
-	sw_rx_data_prod->mapping = replace_buf->mapping;
+	if (bd->page_offset == PAGE_SIZE) {
+		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
+			DP_NOTICE(edev,
+				  "Failed to allocate RX buffer for tpa start\n");
+			bd->page_offset -= rxq->rx_buf_seg_size;
+			page_ref_inc(bd->data);
+			dev_kfree_skb_any(skb);
+			return NULL;
+		}
+	} else {
+		page_ref_inc(bd->data);
+		qede_reuse_page(rxq, bd);
+	}
+
+	/* We've consumed the first BD and prepared an SKB */
+	qede_rx_bd_ring_consume(rxq);
 
-	sw_rx_data_prod->data = replace_buf->data;
-	rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(mapping));
-	rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(mapping));
-	sw_rx_data_prod->page_offset = replace_buf->page_offset;
+	return skb;
+}
 
-	rxq->sw_rx_prod++;
+static struct sk_buff *
+qede_rx_build_skb(struct qede_dev *edev,
+		  struct qede_rx_queue *rxq,
+		  struct sw_rx_data *bd, u16 len, u16 pad)
+{
+	struct sk_buff *skb = NULL;
 
-	/* move partial skb from cons to pool (don't unmap yet)
-	 * save mapping, incase we drop the packet later on.
+	/* For smaller frames still need to allocate skb, memcpy
+	 * data and benefit in reusing the page segment instead of
+	 * un-mapping it.
 	 */
-	tpa_info->buffer = *sw_rx_data_cons;
-	mapping = HILO_U64(le32_to_cpu(rx_bd_cons->addr.hi),
-			   le32_to_cpu(rx_bd_cons->addr.lo));
+	if ((len + pad <= edev->rx_copybreak)) {
+		unsigned int offset = bd->page_offset + pad;
 
-	tpa_info->buffer_mapping = mapping;
-	rxq->sw_rx_cons++;
+		skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
+		if (unlikely(!skb))
+			return NULL;
+
+		skb_reserve(skb, pad);
+		memcpy(skb_put(skb, len),
+		       page_address(bd->data) + offset, len);
+		qede_reuse_page(rxq, bd);
+		goto out;
+	}
 
-	/* set tpa state to start only if we are able to allocate skb
-	 * for this aggregation, otherwise mark as error and aggregation will
-	 * be dropped
+	skb = qede_build_skb(rxq, bd, len, pad);
+
+	if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
+		/* Incr page ref count to reuse on allocation failure so
+		 * that it doesn't get freed while freeing SKB [as its
+		 * already mapped there].
 		 */
-	tpa_info->skb = netdev_alloc_skb(edev->ndev,
-					 le16_to_cpu(cqe->len_on_first_bd));
+		page_ref_inc(bd->data);
+		dev_kfree_skb_any(skb);
+		return NULL;
+	}
+out:
+	/* We've consumed the first BD and prepared an SKB */
+	qede_rx_bd_ring_consume(rxq);
+
+	return skb;
+}
+
+static void qede_tpa_start(struct qede_dev *edev,
+			   struct qede_rx_queue *rxq,
+			   struct eth_fast_path_rx_tpa_start_cqe *cqe)
+{
+	struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
+	struct sw_rx_data *sw_rx_data_cons;
+	u16 pad;
+
+	sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
+	pad = cqe->placement_offset + rxq->rx_headroom;
+
+	tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
+					      le16_to_cpu(cqe->len_on_first_bd),
+					      pad, false);
+	tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
+	tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
+
 	if (unlikely(!tpa_info->skb)) {
 		DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
+
+		/* Consume from ring but do not produce since
+		 * this might be used by FW still, it will be re-used
+		 * at TPA end.
+		 */
+		tpa_info->tpa_start_fail = true;
+		qede_rx_bd_ring_consume(rxq);
 		tpa_info->state = QEDE_AGG_STATE_ERROR;
 		goto cons_buf;
 	}
 
-	/* Start filling in the aggregation info */
-	skb_put(tpa_info->skb, le16_to_cpu(cqe->len_on_first_bd));
 	tpa_info->frag_id = 0;
 	tpa_info->state = QEDE_AGG_STATE_START;
 
-	/* Store some information from first CQE */
-	tpa_info->start_cqe_placement_offset = cqe->placement_offset;
-	tpa_info->start_cqe_bd_len = le16_to_cpu(cqe->len_on_first_bd);
 	if ((le16_to_cpu(cqe->pars_flags.flags) >>
 	     PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
 	    PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
@@ -899,6 +964,10 @@ static int qede_tpa_end(struct qede_dev *edev,
 	tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
 	skb = tpa_info->skb;
 
+	if (tpa_info->buffer.page_offset == PAGE_SIZE)
+		dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
+			       PAGE_SIZE, rxq->data_direction);
+
 	for (i = 0; cqe->len_list[i]; i++)
 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
 				   le16_to_cpu(cqe->len_list[i]));
@@ -919,11 +988,6 @@ static int qede_tpa_end(struct qede_dev *edev,
 		       "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
 		       le16_to_cpu(cqe->total_packet_len), skb->len);
 
-	memcpy(skb->data,
-	       page_address(tpa_info->buffer.data) +
-	       tpa_info->start_cqe_placement_offset +
-	       tpa_info->buffer.page_offset, tpa_info->start_cqe_bd_len);
-
 	/* Finalize the SKB */
 	skb->protocol = eth_type_trans(skb, edev->ndev);
 	skb->ip_summed = CHECKSUM_UNNECESSARY;
@@ -940,6 +1004,12 @@ static int qede_tpa_end(struct qede_dev *edev,
 	return 1;
 err:
 	tpa_info->state = QEDE_AGG_STATE_NONE;
+
+	if (tpa_info->tpa_start_fail) {
+		qede_reuse_page(rxq, &tpa_info->buffer);
+		tpa_info->tpa_start_fail = false;
+	}
+
 	dev_kfree_skb_any(tpa_info->skb);
 	tpa_info->skb = NULL;
 	return 0;
@@ -1058,65 +1128,6 @@ static bool qede_rx_xdp(struct qede_dev *edev,
 	return false;
 }
 
-static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev,
-					    struct qede_rx_queue *rxq,
-					    struct sw_rx_data *bd, u16 len,
-					    u16 pad)
-{
-	unsigned int offset = bd->page_offset + pad;
-	struct skb_frag_struct *frag;
-	struct page *page = bd->data;
-	unsigned int pull_len;
-	struct sk_buff *skb;
-	unsigned char *va;
-
-	/* Allocate a new SKB with a sufficient large header len */
-	skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
-	if (unlikely(!skb))
-		return NULL;
-
-	/* Copy data into SKB - if it's small, we can simply copy it and
-	 * re-use the already allcoated & mapped memory.
-	 */
-	if (len + pad <= edev->rx_copybreak) {
-		skb_put_data(skb, page_address(page) + offset, len);
-		qede_reuse_page(rxq, bd);
-		goto out;
-	}
-
-	frag = &skb_shinfo(skb)->frags[0];
-
-	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
-			page, offset, len, rxq->rx_buf_seg_size);
-
-	va = skb_frag_address(frag);
-	pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE);
-
-	/* Align the pull_len to optimize memcpy */
-	memcpy(skb->data, va, ALIGN(pull_len, sizeof(long)));
-
-	/* Correct the skb & frag sizes offset after the pull */
-	skb_frag_size_sub(frag, pull_len);
-	frag->page_offset += pull_len;
-	skb->data_len -= pull_len;
-	skb->tail += pull_len;
-
-	if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
-		/* Incr page ref count to reuse on allocation failure so
-		 * that it doesn't get freed while freeing SKB [as its
-		 * already mapped there].
-		 */
-		page_ref_inc(page);
-		dev_kfree_skb_any(skb);
-		return NULL;
-	}
-
-out:
-	/* We've consumed the first BD and prepared an SKB */
-	qede_rx_bd_ring_consume(rxq);
-	return skb;
-}
-
 static int qede_rx_build_jumbo(struct qede_dev *edev,
 			       struct qede_rx_queue *rxq,
 			       struct sk_buff *skb,
@@ -1157,7 +1168,7 @@ static int qede_rx_build_jumbo(struct qede_dev *edev,
 			       PAGE_SIZE, DMA_FROM_DEVICE);
 
 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++,
-				   bd->data, 0, cur_size);
+				   bd->data, rxq->rx_headroom, cur_size);
 
 		skb->truesize += PAGE_SIZE;
 		skb->data_len += cur_size;
@@ -1256,7 +1267,7 @@ static int qede_rx_process_cqe(struct qede_dev *edev,
 	/* Basic validation passed; Need to prepare an SKB. This would also
 	 * guarantee to finally consume the first BD upon success.
 	 */
-	skb = qede_rx_allocate_skb(edev, rxq, bd, len, pad);
+	skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
 	if (!skb) {
 		rxq->rx_alloc_errors++;
 		qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);

drivers/net/ethernet/qlogic/qede/qede_main.c

@@ -1197,30 +1197,8 @@ static void qede_free_rx_buffers(struct qede_dev *edev,
 	}
 }
 
-static void qede_free_sge_mem(struct qede_dev *edev, struct qede_rx_queue *rxq)
-{
-	int i;
-
-	if (edev->gro_disable)
-		return;
-
-	for (i = 0; i < ETH_TPA_MAX_AGGS_NUM; i++) {
-		struct qede_agg_info *tpa_info = &rxq->tpa_info[i];
-		struct sw_rx_data *replace_buf = &tpa_info->buffer;
-
-		if (replace_buf->data) {
-			dma_unmap_page(&edev->pdev->dev,
-				       replace_buf->mapping,
-				       PAGE_SIZE, DMA_FROM_DEVICE);
-			__free_page(replace_buf->data);
-		}
-	}
-}
-
 static void qede_free_mem_rxq(struct qede_dev *edev, struct qede_rx_queue *rxq)
 {
-	qede_free_sge_mem(edev, rxq);
-
 	/* Free rx buffers */
 	qede_free_rx_buffers(edev, rxq);
 
@@ -1232,45 +1210,15 @@ static void qede_free_mem_rxq(struct qede_dev *edev, struct qede_rx_queue *rxq)
 	edev->ops->common->chain_free(edev->cdev, &rxq->rx_comp_ring);
 }
 
-static int qede_alloc_sge_mem(struct qede_dev *edev, struct qede_rx_queue *rxq)
+static void qede_set_tpa_param(struct qede_rx_queue *rxq)
 {
-	dma_addr_t mapping;
 	int i;
 
-	if (edev->gro_disable)
-		return 0;
-
 	for (i = 0; i < ETH_TPA_MAX_AGGS_NUM; i++) {
 		struct qede_agg_info *tpa_info = &rxq->tpa_info[i];
-		struct sw_rx_data *replace_buf = &tpa_info->buffer;
 
-		replace_buf->data = alloc_pages(GFP_ATOMIC, 0);
-		if (unlikely(!replace_buf->data)) {
-			DP_NOTICE(edev,
-				  "Failed to allocate TPA skb pool [replacement buffer]\n");
-			goto err;
-		}
-
-		mapping = dma_map_page(&edev->pdev->dev, replace_buf->data, 0,
-				       PAGE_SIZE, DMA_FROM_DEVICE);
-		if (unlikely(dma_mapping_error(&edev->pdev->dev, mapping))) {
-			DP_NOTICE(edev,
-				  "Failed to map TPA replacement buffer\n");
-			goto err;
-		}
-
-		replace_buf->mapping = mapping;
-		tpa_info->buffer.page_offset = 0;
-		tpa_info->buffer_mapping = mapping;
 		tpa_info->state = QEDE_AGG_STATE_NONE;
 	}
-
-	return 0;
-err:
-	qede_free_sge_mem(edev, rxq);
-	edev->gro_disable = 1;
-	edev->ndev->features &= ~NETIF_F_GRO_HW;
-	return -ENOMEM;
 }
 
 /* This function allocates all memory needed per Rx queue */
@@ -1281,19 +1229,24 @@ static int qede_alloc_mem_rxq(struct qede_dev *edev, struct qede_rx_queue *rxq)
 	rxq->num_rx_buffers = edev->q_num_rx_buffers;
 
 	rxq->rx_buf_size = NET_IP_ALIGN + ETH_OVERHEAD + edev->ndev->mtu;
-	rxq->rx_headroom = edev->xdp_prog ? XDP_PACKET_HEADROOM : 0;
+
+	rxq->rx_headroom = edev->xdp_prog ? XDP_PACKET_HEADROOM : NET_SKB_PAD;
+	size = rxq->rx_headroom +
+	       SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 
 	/* Make sure that the headroom and  payload fit in a single page */
-	if (rxq->rx_buf_size + rxq->rx_headroom > PAGE_SIZE)
-		rxq->rx_buf_size = PAGE_SIZE - rxq->rx_headroom;
+	if (rxq->rx_buf_size + size > PAGE_SIZE)
+		rxq->rx_buf_size = PAGE_SIZE - size;
 
-	/* Segment size to spilt a page in multiple equal parts,
+	/* Segment size to spilt a page in multiple equal parts ,
 	 * unless XDP is used in which case we'd use the entire page.
 	 */
-	if (!edev->xdp_prog)
-		rxq->rx_buf_seg_size = roundup_pow_of_two(rxq->rx_buf_size);
-	else
+	if (!edev->xdp_prog) {
+		size = size + rxq->rx_buf_size;
+		rxq->rx_buf_seg_size = roundup_pow_of_two(size);
+	} else {
 		rxq->rx_buf_seg_size = PAGE_SIZE;
+	}
 
 	/* Allocate the parallel driver ring for Rx buffers */
 	size = sizeof(*rxq->sw_rx_ring) * RX_RING_SIZE;
@@ -1337,7 +1290,8 @@ static int qede_alloc_mem_rxq(struct qede_dev *edev, struct qede_rx_queue *rxq)
 		}
 	}
 
-	rc = qede_alloc_sge_mem(edev, rxq);
+	if (!edev->gro_disable)
+		qede_set_tpa_param(rxq);
 err:
 	return rc;
 }