Message-ID: <1480258273-24973-6-git-send-email-Yuval.Mintz@cavium.com>
Date:   Sun, 27 Nov 2016 16:51:07 +0200
From:   Yuval Mintz <Yuval.Mintz@...ium.com>
To:     <davem@...emloft.net>, <netdev@...r.kernel.org>
CC:     Yuval Mintz <Yuval.Mintz@...ium.com>
Subject: [PATCH net-next 05/11] qede: Refactor data-path Rx flow

Driver's NAPI poll is using a long sequence for processing ingress
packets, and it's going to get even longer once we do XDP.
Break down the main loop into a series of sub-functions to allow
better readability of the function.

While we're at it, correct the accounting of the NAPI budget -
currently we're counting only packets passed to the stack against
the budget, even in case those are actually aggregations.
After refactoring every CQE processed would be counted against the budget.

Signed-off-by: Yuval Mintz <Yuval.Mintz@...ium.com>
---
 drivers/net/ethernet/qlogic/qede/qede_main.c | 488 +++++++++++++++------------
 1 file changed, 264 insertions(+), 224 deletions(-)

diff --git a/drivers/net/ethernet/qlogic/qede/qede_main.c b/drivers/net/ethernet/qlogic/qede/qede_main.c
index 2006dd4..acc350b 100644
--- a/drivers/net/ethernet/qlogic/qede/qede_main.c
+++ b/drivers/net/ethernet/qlogic/qede/qede_main.c
@@ -1021,6 +1021,7 @@ static inline void qede_skb_receive(struct qede_dev *edev,
 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
 
 	napi_gro_receive(&fp->napi, skb);
+	fp->rxq->rcv_pkts++;
 }
 
 static void qede_set_gro_params(struct qede_dev *edev,
@@ -1383,238 +1384,298 @@ static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
 	return false;
 }
 
-static int qede_rx_int(struct qede_fastpath *fp, int budget)
+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)
 {
-	struct qede_dev *edev = fp->edev;
-	struct qede_rx_queue *rxq = fp->rxq;
-
-	u16 hw_comp_cons, sw_comp_cons, sw_rx_index, parse_flag;
-	int rx_pkt = 0;
-	u8 csum_flag;
+	unsigned int offset = bd->page_offset;
+	struct skb_frag_struct *frag;
+	struct page *page = bd->data;
+	unsigned int pull_len;
+	struct sk_buff *skb;
+	unsigned char *va;
 
-	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
-	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
+	/* 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;
 
-	/* Memory barrier to prevent the CPU from doing speculative reads of CQE
-	 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
-	 * read before it is written by FW, then FW writes CQE and SB, and then
-	 * the CPU reads the hw_comp_cons, it will use an old CQE.
+	/* Copy data into SKB - if it's small, we can simply copy it and
+	 * re-use the already allcoated & mapped memory.
 	 */
-	rmb();
+	if (len + pad <= edev->rx_copybreak) {
+		memcpy(skb_put(skb, len),
+		       page_address(page) + pad + offset, len);
+		qede_reuse_page(edev, rxq, bd);
+		goto out;
+	}
 
-	/* Loop to complete all indicated BDs */
-	while (sw_comp_cons != hw_comp_cons) {
-		struct eth_fast_path_rx_reg_cqe *fp_cqe;
-		enum pkt_hash_types rxhash_type;
-		enum eth_rx_cqe_type cqe_type;
-		struct sw_rx_data *sw_rx_data;
-		union eth_rx_cqe *cqe;
-		struct sk_buff *skb;
-		struct page *data;
-		__le16 flags;
-		u16 len, pad;
-		u32 rx_hash;
-
-		/* Get the CQE from the completion ring */
-		cqe = (union eth_rx_cqe *)
-			qed_chain_consume(&rxq->rx_comp_ring);
-		cqe_type = cqe->fast_path_regular.type;
-
-		if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
-			edev->ops->eth_cqe_completion(
-					edev->cdev, fp->id,
-					(struct eth_slow_path_rx_cqe *)cqe);
-			goto next_cqe;
-		}
+	frag = &skb_shinfo(skb)->frags[0];
 
-		if (cqe_type != ETH_RX_CQE_TYPE_REGULAR) {
-			switch (cqe_type) {
-			case ETH_RX_CQE_TYPE_TPA_START:
-				qede_tpa_start(edev, rxq,
-					       &cqe->fast_path_tpa_start);
-				goto next_cqe;
-			case ETH_RX_CQE_TYPE_TPA_CONT:
-				qede_tpa_cont(edev, rxq,
-					      &cqe->fast_path_tpa_cont);
-				goto next_cqe;
-			case ETH_RX_CQE_TYPE_TPA_END:
-				qede_tpa_end(edev, fp,
-					     &cqe->fast_path_tpa_end);
-				goto next_rx_only;
-			default:
-				break;
-			}
-		}
+	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
+			page, pad + offset, len, rxq->rx_buf_seg_size);
 
-		/* Get the data from the SW ring */
-		sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
-		sw_rx_data = &rxq->sw_rx_ring[sw_rx_index];
-		data = sw_rx_data->data;
-
-		fp_cqe = &cqe->fast_path_regular;
-		len =  le16_to_cpu(fp_cqe->len_on_first_bd);
-		pad = fp_cqe->placement_offset;
-		flags = cqe->fast_path_regular.pars_flags.flags;
-
-		/* If this is an error packet then drop it */
-		parse_flag = le16_to_cpu(flags);
-
-		csum_flag = qede_check_csum(parse_flag);
-		if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
-			if (qede_pkt_is_ip_fragmented(&cqe->fast_path_regular,
-						      parse_flag)) {
-				rxq->rx_ip_frags++;
-				goto alloc_skb;
-			}
+	va = skb_frag_address(frag);
+	pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE);
 
-			DP_NOTICE(edev,
-				  "CQE in CONS = %u has error, flags = %x, dropping incoming packet\n",
-				  sw_comp_cons, parse_flag);
-			rxq->rx_hw_errors++;
-			qede_recycle_rx_bd_ring(rxq, edev, fp_cqe->bd_num);
-			goto next_cqe;
-		}
+	/* Align the pull_len to optimize memcpy */
+	memcpy(skb->data, va, ALIGN(pull_len, sizeof(long)));
 
-alloc_skb:
-		skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
-		if (unlikely(!skb)) {
-			DP_NOTICE(edev,
-				  "skb allocation failed, dropping incoming packet\n");
-			qede_recycle_rx_bd_ring(rxq, edev, fp_cqe->bd_num);
-			rxq->rx_alloc_errors++;
-			goto next_cqe;
+	/* 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(edev, 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,
+			       struct eth_fast_path_rx_reg_cqe *cqe,
+			       u16 first_bd_len)
+{
+	u16 pkt_len = le16_to_cpu(cqe->pkt_len);
+	struct sw_rx_data *bd;
+	u16 bd_cons_idx;
+	u8 num_frags;
+
+	pkt_len -= first_bd_len;
+
+	/* We've already used one BD for the SKB. Now take care of the rest */
+	for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
+		u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
+		    pkt_len;
+
+		if (unlikely(!cur_size)) {
+			DP_ERR(edev,
+			       "Still got %d BDs for mapping jumbo, but length became 0\n",
+			       num_frags);
+			goto out;
 		}
 
-		/* Copy data into SKB */
-		if (len + pad <= edev->rx_copybreak) {
-			memcpy(skb_put(skb, len),
-			       page_address(data) + pad +
-				sw_rx_data->page_offset, len);
-			qede_reuse_page(edev, rxq, sw_rx_data);
+		/* We need a replacement buffer for each BD */
+		if (unlikely(qede_alloc_rx_buffer(edev, rxq)))
+			goto out;
+
+		/* Now that we've allocated the replacement buffer,
+		 * we can safely consume the next BD and map it to the SKB.
+		 */
+		bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
+		bd = &rxq->sw_rx_ring[bd_cons_idx];
+		qede_rx_bd_ring_consume(rxq);
+
+		dma_unmap_page(&edev->pdev->dev, bd->mapping,
+			       PAGE_SIZE, DMA_FROM_DEVICE);
+
+		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags++,
+				   bd->data, 0, cur_size);
+
+		skb->truesize += PAGE_SIZE;
+		skb->data_len += cur_size;
+		skb->len += cur_size;
+		pkt_len -= cur_size;
+	}
+
+	if (unlikely(pkt_len))
+		DP_ERR(edev,
+		       "Mapped all BDs of jumbo, but still have %d bytes\n",
+		       pkt_len);
+
+out:
+	return num_frags;
+}
+
+static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
+				   struct qede_fastpath *fp,
+				   struct qede_rx_queue *rxq,
+				   union eth_rx_cqe *cqe,
+				   enum eth_rx_cqe_type type)
+{
+	switch (type) {
+	case ETH_RX_CQE_TYPE_TPA_START:
+		qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
+		return 0;
+	case ETH_RX_CQE_TYPE_TPA_CONT:
+		qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
+		return 0;
+	case ETH_RX_CQE_TYPE_TPA_END:
+		qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
+		return 1;
+	default:
+		return 0;
+	}
+}
+
+static int qede_rx_process_cqe(struct qede_dev *edev,
+			       struct qede_fastpath *fp,
+			       struct qede_rx_queue *rxq)
+{
+	struct eth_fast_path_rx_reg_cqe *fp_cqe;
+	u16 len, pad, bd_cons_idx, parse_flag;
+	enum pkt_hash_types rxhash_type;
+	enum eth_rx_cqe_type cqe_type;
+	union eth_rx_cqe *cqe;
+	struct sw_rx_data *bd;
+	struct sk_buff *skb;
+	__le16 flags;
+	u8 csum_flag;
+	u32 rx_hash;
+
+	/* Get the CQE from the completion ring */
+	cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
+	cqe_type = cqe->fast_path_regular.type;
+
+	/* Process an unlikely slowpath event */
+	if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
+		struct eth_slow_path_rx_cqe *sp_cqe;
+
+		sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
+		edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
+		return 0;
+	}
+
+	/* Handle TPA cqes */
+	if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
+		return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
+
+	/* Get the data from the SW ring; Consume it only after its evident
+	 * we wouldn't recycle it.
+	 */
+	bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
+	bd = &rxq->sw_rx_ring[bd_cons_idx];
+
+	fp_cqe = &cqe->fast_path_regular;
+	len = le16_to_cpu(fp_cqe->len_on_first_bd);
+	pad = fp_cqe->placement_offset;
+
+	/* If this is an error packet then drop it */
+	flags = cqe->fast_path_regular.pars_flags.flags;
+	parse_flag = le16_to_cpu(flags);
+
+	csum_flag = qede_check_csum(parse_flag);
+	if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
+		if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag)) {
+			rxq->rx_ip_frags++;
 		} else {
-			struct skb_frag_struct *frag;
-			unsigned int pull_len;
-			unsigned char *va;
-
-			frag = &skb_shinfo(skb)->frags[0];
-
-			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, data,
-					pad + sw_rx_data->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)));
-
-			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(edev, rxq,
-							    sw_rx_data))) {
-				DP_ERR(edev, "Failed to allocate rx buffer\n");
-				/* Incr page ref count to reuse on allocation
-				 * failure so that it doesn't get freed while
-				 * freeing SKB.
-				 */
-
-				page_ref_inc(sw_rx_data->data);
-				rxq->rx_alloc_errors++;
-				qede_recycle_rx_bd_ring(rxq, edev,
-							fp_cqe->bd_num);
-				dev_kfree_skb_any(skb);
-				goto next_cqe;
-			}
+			DP_NOTICE(edev,
+				  "CQE has error, flags = %x, dropping incoming packet\n",
+				  parse_flag);
+			rxq->rx_hw_errors++;
+			qede_recycle_rx_bd_ring(rxq, edev, fp_cqe->bd_num);
+			return 0;
 		}
+	}
 
-		qede_rx_bd_ring_consume(rxq);
+	/* Basic validation passed; Need to prepare an SKB. This would also
+	 * guaranteed to finally consume the first BD upon success.
+	 */
+	skb = qede_rx_allocate_skb(edev, rxq, bd, len, pad);
+	if (!skb) {
+		rxq->rx_alloc_errors++;
+		qede_recycle_rx_bd_ring(rxq, edev, fp_cqe->bd_num);
+		return 0;
+	}
 
-		if (fp_cqe->bd_num != 1) {
-			u16 pkt_len = le16_to_cpu(fp_cqe->pkt_len);
-			u8 num_frags;
-
-			pkt_len -= len;
-
-			for (num_frags = fp_cqe->bd_num - 1; num_frags > 0;
-			     num_frags--) {
-				u16 cur_size = pkt_len > rxq->rx_buf_size ?
-						rxq->rx_buf_size : pkt_len;
-				if (unlikely(!cur_size)) {
-					DP_ERR(edev,
-					       "Still got %d BDs for mapping jumbo, but length became 0\n",
-					       num_frags);
-					qede_recycle_rx_bd_ring(rxq, edev,
-								num_frags);
-					dev_kfree_skb_any(skb);
-					goto next_cqe;
-				}
-
-				if (unlikely(qede_alloc_rx_buffer(edev, rxq))) {
-					qede_recycle_rx_bd_ring(rxq, edev,
-								num_frags);
-					dev_kfree_skb_any(skb);
-					goto next_cqe;
-				}
-
-				sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
-				sw_rx_data = &rxq->sw_rx_ring[sw_rx_index];
-				qede_rx_bd_ring_consume(rxq);
-
-				dma_unmap_page(&edev->pdev->dev,
-					       sw_rx_data->mapping,
-					       PAGE_SIZE, DMA_FROM_DEVICE);
-
-				skb_fill_page_desc(skb,
-						   skb_shinfo(skb)->nr_frags++,
-						   sw_rx_data->data, 0,
-						   cur_size);
-
-				skb->truesize += PAGE_SIZE;
-				skb->data_len += cur_size;
-				skb->len += cur_size;
-				pkt_len -= cur_size;
-			}
+	/* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
+	 * by a single cqe.
+	 */
+	if (fp_cqe->bd_num > 1) {
+		u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
+							 fp_cqe, len);
 
-			if (unlikely(pkt_len))
-				DP_ERR(edev,
-				       "Mapped all BDs of jumbo, but still have %d bytes\n",
-				       pkt_len);
+		if (unlikely(unmapped_frags > 0)) {
+			qede_recycle_rx_bd_ring(rxq, edev, unmapped_frags);
+			dev_kfree_skb_any(skb);
+			return 0;
 		}
+	}
 
-		skb->protocol = eth_type_trans(skb, edev->ndev);
+	/* The SKB contains all the data. Now prepare meta-magic */
+	skb->protocol = eth_type_trans(skb, edev->ndev);
+	rx_hash = qede_get_rxhash(edev, fp_cqe->bitfields,
+				  fp_cqe->rss_hash, &rxhash_type);
+	skb_set_hash(skb, rx_hash, rxhash_type);
+	qede_set_skb_csum(skb, csum_flag);
+	skb_record_rx_queue(skb, rxq->rxq_id);
 
-		rx_hash = qede_get_rxhash(edev, fp_cqe->bitfields,
-					  fp_cqe->rss_hash, &rxhash_type);
+	/* SKB is prepared - pass it to stack */
+	qede_skb_receive(edev, fp, skb, le16_to_cpu(fp_cqe->vlan_tag));
 
-		skb_set_hash(skb, rx_hash, rxhash_type);
+	return 1;
+}
 
-		qede_set_skb_csum(skb, csum_flag);
+static int qede_rx_int(struct qede_fastpath *fp, int budget)
+{
+	struct qede_rx_queue *rxq = fp->rxq;
+	struct qede_dev *edev = fp->edev;
+	u16 hw_comp_cons, sw_comp_cons;
+	int work_done = 0;
 
-		skb_record_rx_queue(skb, fp->rxq->rxq_id);
+	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
+	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
 
-		qede_skb_receive(edev, fp, skb, le16_to_cpu(fp_cqe->vlan_tag));
-next_rx_only:
-		rx_pkt++;
+	/* Memory barrier to prevent the CPU from doing speculative reads of CQE
+	 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
+	 * read before it is written by FW, then FW writes CQE and SB, and then
+	 * the CPU reads the hw_comp_cons, it will use an old CQE.
+	 */
+	rmb();
 
-next_cqe: /* don't consume bd rx buffer */
+	/* Loop to complete all indicated BDs */
+	while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
+		qede_rx_process_cqe(edev, fp, rxq);
 		qed_chain_recycle_consumed(&rxq->rx_comp_ring);
 		sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
-		/* CR TPA - revisit how to handle budget in TPA perhaps
-		 * increase on "end"
-		 */
-		if (rx_pkt == budget)
-			break;
-	} /* repeat while sw_comp_cons != hw_comp_cons... */
+		work_done++;
+	}
 
 	/* Update producers */
 	qede_update_rx_prod(edev, rxq);
 
-	rxq->rcv_pkts += rx_pkt;
+	return work_done;
+}
+
+static bool qede_poll_is_more_work(struct qede_fastpath *fp)
+{
+	qed_sb_update_sb_idx(fp->sb_info);
 
-	return rx_pkt;
+	/* *_has_*_work() reads the status block, thus we need to ensure that
+	 * status block indices have been actually read (qed_sb_update_sb_idx)
+	 * prior to this check (*_has_*_work) so that we won't write the
+	 * "newer" value of the status block to HW (if there was a DMA right
+	 * after qede_has_rx_work and if there is no rmb, the memory reading
+	 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
+	 * In this case there will never be another interrupt until there is
+	 * another update of the status block, while there is still unhandled
+	 * work.
+	 */
+	rmb();
+
+	if (likely(fp->type & QEDE_FASTPATH_RX))
+		if (qede_has_rx_work(fp->rxq))
+			return true;
+
+	if (likely(fp->type & QEDE_FASTPATH_TX))
+		if (qede_txq_has_work(fp->txq))
+			return true;
+
+	return false;
 }
 
 static int qede_poll(struct napi_struct *napi, int budget)
@@ -1631,32 +1692,11 @@ static int qede_poll(struct napi_struct *napi, int budget)
 			qede_has_rx_work(fp->rxq)) ?
 			qede_rx_int(fp, budget) : 0;
 	if (rx_work_done < budget) {
-		qed_sb_update_sb_idx(fp->sb_info);
-		/* *_has_*_work() reads the status block,
-		 * thus we need to ensure that status block indices
-		 * have been actually read (qed_sb_update_sb_idx)
-		 * prior to this check (*_has_*_work) so that
-		 * we won't write the "newer" value of the status block
-		 * to HW (if there was a DMA right after
-		 * qede_has_rx_work and if there is no rmb, the memory
-		 * reading (qed_sb_update_sb_idx) may be postponed
-		 * to right before *_ack_sb). In this case there
-		 * will never be another interrupt until there is
-		 * another update of the status block, while there
-		 * is still unhandled work.
-		 */
-		rmb();
-
-		/* Fall out from the NAPI loop if needed */
-		if (!((likely(fp->type & QEDE_FASTPATH_RX) &&
-		       qede_has_rx_work(fp->rxq)) ||
-		      (likely(fp->type & QEDE_FASTPATH_TX) &&
-		       qede_txq_has_work(fp->txq)))) {
+		if (!qede_poll_is_more_work(fp)) {
 			napi_complete(napi);
 
 			/* Update and reenable interrupts */
-			qed_sb_ack(fp->sb_info, IGU_INT_ENABLE,
-				   1 /*update*/);
+			qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
 		} else {
 			rx_work_done = budget;
 		}
-- 
1.9.3

Powered by blists - more mailing lists