net: hns3: Add handling of GRO Pkts not fully RX'ed in NAPI poll

The "FE bit" in the description means the last description for
a packets. When HW GRO enable, HW write data to ring every
packet/buffer, there is greater probability that driver handle
with the describtion but HW still not set the "FE bit".

When drier handle the packet and HW still not set "FE bit",
driver stores skb and bd_num in rx ring, and continue to use the
skb and bd_num in next napi.
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/hisilicon/hns3/hns3_enet.c

@@ -2383,6 +2383,90 @@ static bool hns3_parse_vlan_tag(struct hns3_enet_ring *ring,
 	}
 }
 
+static int hns3_alloc_skb(struct hns3_enet_ring *ring, int length,
+			  unsigned char *va)
+{
+#define HNS3_NEED_ADD_FRAG	1
+	struct hns3_desc_cb *desc_cb = &ring->desc_cb[ring->next_to_clean];
+	struct net_device *netdev = ring->tqp->handle->kinfo.netdev;
+	struct sk_buff *skb;
+
+	ring->skb = napi_alloc_skb(&ring->tqp_vector->napi, HNS3_RX_HEAD_SIZE);
+	skb = ring->skb;
+	if (unlikely(!skb)) {
+		netdev_err(netdev, "alloc rx skb fail\n");
+
+		u64_stats_update_begin(&ring->syncp);
+		ring->stats.sw_err_cnt++;
+		u64_stats_update_end(&ring->syncp);
+
+		return -ENOMEM;
+	}
+
+	prefetchw(skb->data);
+
+	ring->pending_buf = 1;
+	if (length <= HNS3_RX_HEAD_SIZE) {
+		memcpy(__skb_put(skb, length), va, ALIGN(length, sizeof(long)));
+
+		/* We can reuse buffer as-is, just make sure it is local */
+		if (likely(page_to_nid(desc_cb->priv) == numa_node_id()))
+			desc_cb->reuse_flag = 1;
+		else /* This page cannot be reused so discard it */
+			put_page(desc_cb->priv);
+
+		ring_ptr_move_fw(ring, next_to_clean);
+		return 0;
+	}
+	u64_stats_update_begin(&ring->syncp);
+	ring->stats.seg_pkt_cnt++;
+	u64_stats_update_end(&ring->syncp);
+
+	ring->pull_len = eth_get_headlen(va, HNS3_RX_HEAD_SIZE);
+	__skb_put(skb, ring->pull_len);
+	hns3_nic_reuse_page(skb, 0, ring, ring->pull_len,
+			    desc_cb);
+	ring_ptr_move_fw(ring, next_to_clean);
+
+	return HNS3_NEED_ADD_FRAG;
+}
+
+static int hns3_add_frag(struct hns3_enet_ring *ring, struct hns3_desc *desc,
+			 struct sk_buff **out_skb, bool pending)
+{
+	struct sk_buff *skb = *out_skb;
+	struct hns3_desc_cb *desc_cb;
+	struct hns3_desc *pre_desc;
+	u32 bd_base_info;
+	int pre_bd;
+
+	/* if there is pending bd, the SW param next_to_clean has moved
+	 * to next and the next is NULL
+	 */
+	if (pending) {
+		pre_bd = (ring->next_to_clean - 1 + ring->desc_num) %
+			ring->desc_num;
+		pre_desc = &ring->desc[pre_bd];
+		bd_base_info = le32_to_cpu(pre_desc->rx.bd_base_info);
+	} else {
+		bd_base_info = le32_to_cpu(desc->rx.bd_base_info);
+	}
+
+	while (!hnae3_get_bit(bd_base_info, HNS3_RXD_FE_B)) {
+		desc = &ring->desc[ring->next_to_clean];
+		desc_cb = &ring->desc_cb[ring->next_to_clean];
+		bd_base_info = le32_to_cpu(desc->rx.bd_base_info);
+		if (!hnae3_get_bit(bd_base_info, HNS3_RXD_VLD_B))
+			return -ENXIO;
+
+		hns3_nic_reuse_page(skb, ring->pending_buf, ring, 0, desc_cb);
+		ring_ptr_move_fw(ring, next_to_clean);
+		ring->pending_buf++;
+	}
+
+	return 0;
+}
+
 static void hns3_set_rx_skb_rss_type(struct hns3_enet_ring *ring,
 				     struct sk_buff *skb)
 {
@@ -2399,18 +2483,16 @@ static void hns3_set_rx_skb_rss_type(struct hns3_enet_ring *ring,
 }
 
 static int hns3_handle_rx_bd(struct hns3_enet_ring *ring,
-			     struct sk_buff **out_skb, int *out_bnum)
+			     struct sk_buff **out_skb)
 {
 	struct net_device *netdev = ring->tqp->handle->kinfo.netdev;
+	struct sk_buff *skb = ring->skb;
 	struct hns3_desc_cb *desc_cb;
 	struct hns3_desc *desc;
-	struct sk_buff *skb;
-	unsigned char *va;
 	u32 bd_base_info;
-	int pull_len;
 	u32 l234info;
 	int length;
-	int bnum;
+	int ret;
 
 	desc = &ring->desc[ring->next_to_clean];
 	desc_cb = &ring->desc_cb[ring->next_to_clean];
@@ -2422,9 +2504,10 @@ static int hns3_handle_rx_bd(struct hns3_enet_ring *ring,
 
 	/* Check valid BD */
 	if (unlikely(!hnae3_get_bit(bd_base_info, HNS3_RXD_VLD_B)))
-		return -EFAULT;
+		return -ENXIO;
 
-	va = (unsigned char *)desc_cb->buf + desc_cb->page_offset;
+	if (!skb)
+		ring->va = (unsigned char *)desc_cb->buf + desc_cb->page_offset;
 
 	/* Prefetch first cache line of first page
 	 * Idea is to cache few bytes of the header of the packet. Our L1 Cache
@@ -2433,62 +2516,42 @@ static int hns3_handle_rx_bd(struct hns3_enet_ring *ring,
 	 * lines. In such a case, single fetch would suffice to cache in the
 	 * relevant part of the header.
 	 */
-	prefetch(va);
+	prefetch(ring->va);
 #if L1_CACHE_BYTES < 128
-	prefetch(va + L1_CACHE_BYTES);
+	prefetch(ring->va + L1_CACHE_BYTES);
 #endif
 
-	skb = *out_skb = napi_alloc_skb(&ring->tqp_vector->napi,
-					HNS3_RX_HEAD_SIZE);
-	if (unlikely(!skb)) {
-		netdev_err(netdev, "alloc rx skb fail\n");
+	if (!skb) {
+		ret = hns3_alloc_skb(ring, length, ring->va);
+		*out_skb = skb = ring->skb;
 
-		u64_stats_update_begin(&ring->syncp);
-		ring->stats.sw_err_cnt++;
-		u64_stats_update_end(&ring->syncp);
-
-		return -ENOMEM;
-	}
-
-	prefetchw(skb->data);
-
-	bnum = 1;
-	if (length <= HNS3_RX_HEAD_SIZE) {
-		memcpy(__skb_put(skb, length), va, ALIGN(length, sizeof(long)));
-
-		/* We can reuse buffer as-is, just make sure it is local */
-		if (likely(page_to_nid(desc_cb->priv) == numa_node_id()))
-			desc_cb->reuse_flag = 1;
-		else /* This page cannot be reused so discard it */
-			put_page(desc_cb->priv);
+		if (ret < 0) /* alloc buffer fail */
+			return ret;
+		if (ret > 0) { /* need add frag */
+			ret = hns3_add_frag(ring, desc, &skb, false);
+			if (ret)
+				return ret;
 
-		ring_ptr_move_fw(ring, next_to_clean);
+			/* As the head data may be changed when GRO enable, copy
+			 * the head data in after other data rx completed
+			 */
+			memcpy(skb->data, ring->va,
+			       ALIGN(ring->pull_len, sizeof(long)));
+		}
 	} else {
-		u64_stats_update_begin(&ring->syncp);
-		ring->stats.seg_pkt_cnt++;
-		u64_stats_update_end(&ring->syncp);
-
-		pull_len = eth_get_headlen(va, HNS3_RX_HEAD_SIZE);
-
-		memcpy(__skb_put(skb, pull_len), va,
-		       ALIGN(pull_len, sizeof(long)));
-
-		hns3_nic_reuse_page(skb, 0, ring, pull_len, desc_cb);
-		ring_ptr_move_fw(ring, next_to_clean);
+		ret = hns3_add_frag(ring, desc, &skb, true);
+		if (ret)
+			return ret;
 
-		while (!hnae3_get_bit(bd_base_info, HNS3_RXD_FE_B)) {
-			desc = &ring->desc[ring->next_to_clean];
-			desc_cb = &ring->desc_cb[ring->next_to_clean];
-			bd_base_info = le32_to_cpu(desc->rx.bd_base_info);
-			hns3_nic_reuse_page(skb, bnum, ring, 0, desc_cb);
-			ring_ptr_move_fw(ring, next_to_clean);
-			bnum++;
-		}
+		/* As the head data may be changed when GRO enable, copy
+		 * the head data in after other data rx completed
+		 */
+		memcpy(skb->data, ring->va,
+		       ALIGN(ring->pull_len, sizeof(long)));
 	}
 
-	*out_bnum = bnum;
-
 	l234info = le32_to_cpu(desc->rx.l234_info);
+	bd_base_info = le32_to_cpu(desc->rx.bd_base_info);
 
 	/* Based on hw strategy, the tag offloaded will be stored at
 	 * ot_vlan_tag in two layer tag case, and stored at vlan_tag
@@ -2539,6 +2602,7 @@ static int hns3_handle_rx_bd(struct hns3_enet_ring *ring,
 	ring->tqp_vector->rx_group.total_bytes += skb->len;
 
 	hns3_rx_checksum(ring, skb, desc);
+	*out_skb = skb;
 	hns3_set_rx_skb_rss_type(ring, skb);
 
 	return 0;
@@ -2551,9 +2615,9 @@ int hns3_clean_rx_ring(
 #define RCB_NOF_ALLOC_RX_BUFF_ONCE 16
 	struct net_device *netdev = ring->tqp->handle->kinfo.netdev;
 	int recv_pkts, recv_bds, clean_count, err;
-	int unused_count = hns3_desc_unused(ring);
-	struct sk_buff *skb = NULL;
-	int num, bnum = 0;
+	int unused_count = hns3_desc_unused(ring) - ring->pending_buf;
+	struct sk_buff *skb = ring->skb;
+	int num;
 
 	num = readl_relaxed(ring->tqp->io_base + HNS3_RING_RX_RING_FBDNUM_REG);
 	rmb(); /* Make sure num taken effect before the other data is touched */
@@ -2567,24 +2631,32 @@ int hns3_clean_rx_ring(
 			hns3_nic_alloc_rx_buffers(ring,
 						  clean_count + unused_count);
 			clean_count = 0;
-			unused_count = hns3_desc_unused(ring);
+			unused_count = hns3_desc_unused(ring) -
+					ring->pending_buf;
 		}
 
 		/* Poll one pkt */
-		err = hns3_handle_rx_bd(ring, &skb, &bnum);
+		err = hns3_handle_rx_bd(ring, &skb);
 		if (unlikely(!skb)) /* This fault cannot be repaired */
 			goto out;
 
-		recv_bds += bnum;
-		clean_count += bnum;
-		if (unlikely(err)) {  /* Do jump the err */
-			recv_pkts++;
+		if (err == -ENXIO) { /* Do not get FE for the packet */
+			goto out;
+		} else if (unlikely(err)) {  /* Do jump the err */
+			recv_bds += ring->pending_buf;
+			clean_count += ring->pending_buf;
+			ring->skb = NULL;
+			ring->pending_buf = 0;
 			continue;
 		}
 
 		/* Do update ip stack process */
 		skb->protocol = eth_type_trans(skb, netdev);
 		rx_fn(ring, skb);
+		recv_bds += ring->pending_buf;
+		clean_count += ring->pending_buf;
+		ring->skb = NULL;
+		ring->pending_buf = 0;
 
 		recv_pkts++;
 	}

drivers/net/ethernet/hisilicon/hns3/hns3_enet.h

@@ -401,11 +401,17 @@ struct hns3_enet_ring {
 	 */
 	int next_to_clean;
 
+	int pull_len; /* head length for current packet */
+	unsigned char *va; /* first buffer address for current packet */
+
 	u32 flag;          /* ring attribute */
 	int irq_init_flag;
 
 	int numa_node;
 	cpumask_t affinity_mask;
+
+	int pending_buf;
+	struct sk_buff *skb;
 };
 
 struct hns_queue;