net: hand off skb list to other cpu to submit to upper layer
Subsystems (2)
RFC PATCH v2 3/3 net: hand off skb list to other cpu to submit to upper layer
The 3rd patch is totally example on ixgbe driver. Pls. doesn't apply it. It's just for a demo.
NIC driver could use this capability like below steps:
1) Call alloc_etherdev_rxtx_mq when probe the NIC to initiate;
In NAPI RX cleanup function:
2) Initiate a local var struct sk_buff_head skb_head;
3) get the cpu number by calling netif_rx_processing_cpu;
4) In the packet collection loop, just calls __skb_queue_tail(skb_head, skb) to add skb to the list;
5) Before exiting, calls raise_netif_irq to submit the skb list to the specific cpu.
We can add another step after 3) to check input_pkt_alien_queue.qlen. If qlen is bigger than netdev_max_backlog, exit the function instead of collecting packets, so NIC hardware could drop packets. It's better than dropping packets by software.
Below _SAMPLE_ patch (has some garbage codes) is against the latest IXGBE driver.
---
src/ixgbe_main.c +33 -17
@@ -443,7 +443,8 @@ static int __ixgbe_notify_dca(struct devstatic void ixgbe_receive_skb (struct ixgbe_adapter *adapter,struct sk_buff *skb,u8 status,struct ixgbe_ring *ring, -union ixgbe_adv_rx_desc *rx_desc) +union ixgbe_adv_rx_desc *rx_desc, +struct sk_buff_head *skb_head) {int ret;bool is_vlan = (status &IXGBE_RXD_STAT_VP );
@@ -469,7 +470,9 @@ static void ixgbe_receive_skb(struct ixgif (adapter->vlgrp && is_vlan && (tag !=0 ))vlan_hwaccel_receive_skb (skb, adapter->vlgrp, tag);else -netif_receive_skb (skb); +__skb_queue_tail (skb_head, skb); +//netif_rx_queue(skb, skb_head); +//YMZHANG netif_receive_skb(skb); #else netif_receive_skb (skb);#endif
@@ -664,7 +667,8 @@ static void ixgbe_lro_ring_flush(structstruct ixgbe_adapter *adapter, struct ixgbe_lro_desc *lrod, u8 status, struct ixgbe_ring *rx_ring, - union ixgbe_adv_rx_desc *rx_desc) + union ixgbe_adv_rx_desc *rx_desc, + struct sk_buff_head *skb_head) { struct iphdr *iph;struct tcphdr *th;
@@ -701,7 +705,7 @@ static void ixgbe_lro_ring_flush(struct#ifdef NETIF_F_TSO skb_shinfo (skb)->gso_size = lrod->mss;#endif -ixgbe_receive_skb (adapter, skb, status, rx_ring, rx_desc); +ixgbe_receive_skb (adapter, skb, status, rx_ring, rx_desc, skb_head); netdev->last_rx = jiffies; lro_data->stats.coal += lrod->append_cnt +1 ;
@@ -718,14 +722,15 @@ static void ixgbe_lro_ring_flush(structstatic void ixgbe_lro_ring_flush_all (struct ixgbe_lro_list *lrolist,struct ixgbe_adapter *adapter,u8 status,struct ixgbe_ring *rx_ring, -union ixgbe_adv_rx_desc *rx_desc) +union ixgbe_adv_rx_desc *rx_desc, +struct sk_buff_head *skb_head) {struct ixgbe_lro_desc *lrod;struct hlist_node *node, *node2;hlist_for_each_entry_safe (lrod, node, node2, &lrolist->active, lro_node)ixgbe_lro_ring_flush (lrolist, adapter, lrod, status, rx_ring, - rx_desc); + rx_desc, skb_head); }/*
@@ -855,14 +860,14 @@ static int ixgbe_lro_ring_queue(struct iif (!header_ok) {ixgbe_lro_ring_flush (lrolist, adapter, lrod, - status, rx_ring, rx_desc); + status, rx_ring, rx_desc, skb_head);return -1 ; }if (seq != lrod->next_seq) {/* out of order packet */ ixgbe_lro_ring_flush (lrolist, adapter, lrod, - status, rx_ring, rx_desc); + status, rx_ring, rx_desc, skb_head);return -1 ; }
@@ -872,7 +877,7 @@ static int ixgbe_lro_ring_queue(struct iif (lrod->tsval > tsval || *(ts_ptr +2 ) ==0 ) {ixgbe_lro_ring_flush (lrolist, adapter, lrod, status, - rx_ring, rx_desc); + rx_ring, rx_desc, skb_head);return -1 ; } lrod->tsval = tsval;
@@ -911,13 +916,13 @@ static int ixgbe_lro_ring_queue(struct i (struct tcphdr *)(lro_skb->data +sizeof (*iph)); header_th->psh |= th->psh;ixgbe_lro_ring_flush (lrolist, adapter, lrod, - status, rx_ring, rx_desc); + status, rx_ring, rx_desc, skb_head);return 0 ; }if (lrod->append_cnt >= lro_data->max)ixgbe_lro_ring_flush (lrolist, adapter, lrod, - status, rx_ring, rx_desc); + status, rx_ring, rx_desc, skb_head);return 0 ; }/*End of if*/
@@ -1001,13 +1006,14 @@ static void ixgbe_lro_ring_init(struct i#endif /* IXGBE_NO_LRO */ +#ifdef CONFIG_IXGBE_NAPI static bool ixgbe_clean_rx_irq (struct ixgbe_adapter *adapter, -struct ixgbe_ring *rx_ring, - int *work_done, int work_to_do) +struct ixgbe_ring *rx_ring, + int *work_done, int work_to_do)#else static bool ixgbe_clean_rx_irq (struct ixgbe_adapter *adapter, - struct ixgbe_ring *rx_ring) + struct ixgbe_ring *rx_ring)#endif {struct pci_dev *pdev = adapter->pdev;
@@ -1019,12 +1025,17 @@ static bool ixgbe_clean_rx_irq(struct ixu16 hdr_info;bool cleaned = false;int cleaned_count =0 ; +struct sk_buff_head skb_head; +int cpu =netif_rx_processing_cpu (adapter->netdev, rx_ring->queue_index); +#ifndef CONFIG_IXGBE_NAPI int work_to_do = rx_ring->work_limit, local_work_done =0 ;int *work_done = &local_work_done;#endif unsigned int total_rx_bytes =0 , total_rx_packets =0 ; +skb_queue_head_init (&skb_head); + i = rx_ring->next_to_clean; rx_desc =IXGBE_RX_DESC_ADV (*rx_ring, i); staterr =le32_to_cpu (rx_desc->wb.upper.status_error);
@@ -1135,7 +1146,7 @@ static bool ixgbe_clean_rx_irq(struct ix goto next_desc; }#endif -ixgbe_receive_skb (adapter, skb, staterr, rx_ring, rx_desc); +ixgbe_receive_skb (adapter, skb, staterr, rx_ring, rx_desc, &skb_head); adapter->netdev->last_rx = jiffies; next_desc:
@@ -1157,7 +1168,7 @@ next_desc: rx_ring->next_to_clean = i;#ifndef IXGBE_NO_LRO ixgbe_lro_ring_flush_all (rx_ring->lrolist, adapter, - staterr, rx_ring, rx_desc); + staterr, rx_ring, rx_desc, skb_head);#endif /* IXGBE_NO_LRO */ cleaned_count =IXGBE_DESC_UNUSED (rx_ring);#ifndef IXGBE_NO_INET_LRO
@@ -1180,6 +1191,9 @@ next_desc:if (*work_done >= work_to_do)IXGBE_WRITE_REG (&adapter->hw,IXGBE_EICS , rx_ring->v_idx);#endif + +raise_netif_irq (cpu, &skb_head); +return cleaned; }
@@ -4103,6 +4117,8 @@ void ixgbe_napi_add_all(struct ixgbe_adafor (q_idx =0 ; q_idx < q_vectors; q_idx++) {struct ixgbe_q_vector *q_vector = &adapter->q_vector[q_idx];netif_napi_add (adapter->netdev, &q_vector->napi, (*poll),64 ); +/*YMZ*/ +//netif_napi_add(adapter->netdev, &q_vector->napi, (*poll), 32); } }
@@ -4998,7 +5014,7 @@ static int __devinit ixgbe_probe(structpci_set_master (pdev);#ifdef HAVE_TX_MQ - netdev =alloc_etherdev_mq (sizeof (struct ixgbe_adapter ),MAX_TX_QUEUES ); + netdev =alloc_etherdev_rxtx_mq (sizeof (struct ixgbe_adapter ),MAX_RX_QUEUES ,MAX_TX_QUEUES );#else netdev =alloc_etherdev (sizeof (struct ixgbe_adapter ));#endif