e100网卡驱动分析
时间：2010-10-15 来源：sandflee
e100网卡驱动
在模块的初始化函数中调用pci_register_driver(&e100_driver)向内核注册e100驱动
2911 static struct pci_driver e100_driver = {
2912 .name = DRV_NAME,
2913 .id_table = e100_id_table,
2914 .probe = e100_probe,
2915 .remove = __devexit_p(e100_remove),
2916 #ifdef CONFIG_PM
2917 /* Power Management hooks */
2918 .suspend = e100_suspend,
2919 .resume = e100_resume,
2920 #endif
2921 .shutdown = e100_shutdown,
2922 .err_handler = &e100_err_handler,
2923 };
内核检测到网卡后（vid,pid在idtable中），会调用e100_probe函数，probe函数创建netdevice结构和e100私有的nic结构，将其初始化，并通过register_netdev注册对应的netdevice结构，具体流程如下：
//为其分配内存
netdev = alloc_etherdev(sizeof(struct nic))
//初始化对应的函数指针 2627 netdev->open = e100_open; 2628 netdev->stop = e100_close; 2629 netdev->hard_start_xmit = e100_xmit_frame; 2630 netdev->set_multicast_list = e100_set_multicast_list; 2631 netdev->set_mac_address = e100_set_mac_address; 2632 netdev->change_mtu = e100_change_mtu; 2633 netdev->do_ioctl = e100_do_ioctl; 2634 SET_ETHTOOL_OPS(netdev, &e100_ethtool_ops); 2635 netdev->tx_timeout = e100_tx_timeout; 2636 netdev->watchdog_timeo = E100_WATCHDOG_PERIOD;
//napi支持，分配napi结构，初始化，并挂在netdevice的napi_list上 2642 nic = netdev_priv(netdev); 2643 netif_napi_add(netdev, &nic->napi, e100_poll, E100_NAPI_WEIGHT);
//nic的初始化 2702 init_timer(&nic->watchdog); 2703 nic->watchdog.function = e100_watchdog; 2704 nic->watchdog.data = (unsigned long)nic; 2705 init_timer(&nic->blink_timer); 2706 nic->blink_timer.function = e100_blink_led; 2707 nic->blink_timer.data = (unsigned long)nic; 2708 2709 INIT_WORK(&nic->tx_timeout_task, e100_tx_timeout_task);
//向内核注册 register_netdev(netdev)
当驱动激活的时候调用e100_open()函数 //初始化和分配rxs数据接收链表 2124 if((err = e100_rx_alloc_list(nic))) 2125 return err; //分配和初始化cbs命令链表，数据的发送通过这个 2126 if((err = e100_alloc_cbs(nic))) 2127 goto err_rx_clean_list; 2128 if((err = e100_hw_init(nic))) 2129 goto err_clean_cbs; 2130 e100_set_multicast_list(nic->netdev); 2131 e100_start_receiver(nic, NULL); 2132 mod_timer(&nic->watchdog, jiffies); 2133 if((err = request_irq(nic->pdev->irq, e100_intr, IRQF_SHARED, 2134 nic->netdev->name, nic->netdev))) 2135 goto err_no_irq; 2136 netif_wake_queue(nic->netdev); 2137 napi_enable(&nic->napi); 数据的发送 static int e100_xmit_frame(struct sk_buff *skb, struct net_device *netdev) 1637 { 1650 err = e100_exec_cb(nic, skb, e100_xmit_prepare); 1651 1652 switch(err) { 1653 case -ENOSPC: 1654 /* We queued the skb, but now we're out of space. */ 1655 DPRINTK(TX_ERR, DEBUG, "No space for CB\n"); 1656 netif_stop_queue(netdev); 1657 break; 1658 case -ENOMEM: 1659 /* This is a hard error - log it. */ 1660 DPRINTK(TX_ERR, DEBUG, "Out of Tx resources, returning skb\n"); 1661 netif_stop_queue(netdev); 1662 return 1; 1663 } 1664 1665 netdev->trans_start = jiffies; 1666 return 0; 1667 } 通过e100_exec_cb()发送数据，并对发送的错误进行处理，如果发送失败，并net_device的发送队列设置相应的标志，参见netif_stop_queue。 其中 e100_xmit_prepare对cb作一些设置，e100_exec_cb才发送数据. 1619 static void e100_xmit_prepare(struct nic *nic, struct cb *cb, 1620 struct sk_buff *skb) 1621 { 1622 cb->command = nic->tx_command; 1623 /* interrupt every 16 packets regardless of delay */ 1624 if((nic->cbs_avail & ~15) == nic->cbs_avail) 1625 cb->command |= cpu_to_le16(cb_i); 1626 cb->u.tcb.tbd_array = cb->dma_addr + offsetof(struct cb, u.tcb.tbd); 1627 cb->u.tcb.tcb_byte_count = 0; 1628 cb->u.tcb.threshold = nic->tx_threshold; 1629 cb->u.tcb.tbd_count = 1; 1630 cb->u.tcb.tbd.buf_addr = cpu_to_le32(pci_map_single(nic->pdev, 1631 skb->data, skb->len, PCI_DMA_TODEVICE)); 1632 /* check for mapping failure? */ 1633 cb->u.tcb.tbd.size = cpu_to_le16(skb->len); 1634 }
e100_exec_cb()则对nic上的cb链表进行处理 837 static int e100_exec_cb(struct nic *nic, struct sk_buff *skb, 838 void (*cb_prepare)(struct nic *, struct cb *, struct sk_buff *)) 839 { //使用一个空闲cb结构 851 cb = nic->cb_to_use; 852 nic->cb_to_use = cb->next; 853 nic->cbs_avail--; 854 cb->skb = skb; //调用e100_xmit_prepare对cb进行设置 859 cb_prepare(nic, cb, skb);
//对链表上的cb进行处理 867 while(nic->cb_to_send != nic->cb_to_use) { 868 if(unlikely(e100_exec_cmd(nic, nic->cuc_cmd, 869 nic->cb_to_send->dma_addr))) { 870 /* Ok, here's where things get sticky. It's 871 * possible that we can't schedule the command 872 * because the controller is too busy, so 873 * let's just queue the command and try again 874 * when another command is scheduled. */ 875 if(err == -ENOSPC) { 876 //request a reset 877 schedule_work(&nic->tx_timeout_task); 878 } 879 break; 880 } else { 881 nic->cuc_cmd = cuc_resume; 882 nic->cb_to_send = nic->cb_to_send->next; 883 } 884 }
e100_exec_cmd执行具体的操作，到这儿，哥就不懂了，有很多硬件相关的东西。有那些cmd?dma起了什么作用？
数据的接受 注册的中断函数 e100_intr，经典的中断处理例程 2032 static irqreturn_t e100_intr(int irq, void *dev_id) 2033 { 2034 struct net_device *netdev = dev_id; 2035 struct nic *nic = netdev_priv(netdev); //读取寄存器，判断中断是我们产生的还是其他设备产生的 2036 u8 stat_ack = ioread8(&nic->csr->scb.stat_ack); 2037 2038 DPRINTK(INTR, DEBUG, "stat_ack = 0x%02X\n", stat_ack); 2039 2040 if(stat_ack == stat_ack_not_ours || /* Not our interrupt */ 2041 stat_ack == stat_ack_not_present) /* Hardware is ejected */ 2042 return IRQ_NONE; 2043 //ack 我们的设备 2044 /* Ack interrupt(s) */ 2045 iowrite8(stat_ack, &nic->csr->scb.stat_ack); 2046 2047 /* We hit Receive No Resource (RNR); restart RU after cleaning */ 2048 if(stat_ack & stat_ack_rnr) 2049 nic->ru_running = RU_SUSPENDED; 2050 //把设备的poll放到softnet_data的poll_list上 2051 if(likely(netif_rx_schedule_prep(netdev, &nic->napi))) { 2052 e100_disable_irq(nic); 2053 __netif_rx_schedule(netdev, &nic->napi); 2054 } 2055 2056 return IRQ_HANDLED; 2057 } 数据的接收用的是napi方法，netif_rx_schedule()最终会调用napi_schedule把设备的poll方法挂在poll_list上。 2359 void __napi_schedule(struct napi_struct *n) 2360 { 2361 unsigned long flags; 2362 2363 local_irq_save(flags); 2364 list_add_tail(&n->poll_list, &__get_cpu_var(softnet_data).poll_list); 2365 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 2366 local_irq_restore(flags); 2367 } 设备的poll方法 2059 static int e100_poll(struct napi_struct *napi, int budget) 2060 { 2061 struct nic *nic = container_of(napi, struct nic, napi); 2062 struct net_device *netdev = nic->netdev; 2063 unsigned int work_done = 0; 2064 2065 e100_rx_clean(nic, &work_done, budget); 2066 e100_tx_clean(nic); 2067 2068 /* If budget not fully consumed, exit the polling mode */ 2069 if (work_done < budget) { 2070 netif_rx_complete(netdev, napi); 2071 e100_enable_irq(nic); 2072 } 2073 2074 return work_done; 2075 } 最终调用netif_recive_skb将数据包提交到上层。具体到硬件还是不太懂啊.：（
内核对数据的接收 有两种模式中断和NAPI。 先说核心的数据结构 996 /* 997 * Incoming packets are placed on per-cpu queues so that 998 * no locking is needed. 999 */ 1000 struct softnet_data 1001 { 1002 struct Qdisc *output_queue; 1003 struct sk_buff_head input_pkt_queue; 1004 struct list_head poll_list; 1005 struct sk_buff *completion_queue; 1006 1007 struct napi_struct backlog; 1008 #ifdef CONFIG_NET_DMA 1009 struct dma_chan *net_dma; 1010 #endif 1011 }; 正如注释所说，每个CPU都有一个，当涉及它的操作时，先关中断，所以不需要锁的保护 普通中断模式下，中断通过netif_rx把数据提交给input_pkt_queue,如果backlog没有添加到poll_list中，添加之。 而在NAPI下只是调用netif_rx_schedule判断是否需要把设备的poll方法添加到poll_list中. 添加后raise_softirq(NET_RX_ACTION),调用net_rx_action软中断进行后续的处理； net_rx_action遍历poll_list链表，调用其poll方法，对非NAPI设备来说就是通过process_backlog通用的poll方法把input_pkt_queue上的数据提交到上层，而对NAPI来说调用自身的poll方法。其中包含流量控制和对poll调用次数的限制。 process_backlog通过netif_receive_skb处理桥接，数据包监听，最后ip_rcv传给ip层 以e100_poll为例，其遍历rxs接收数据链表，对缓冲区的数据作unmap操作，然后调用netif_receive_skb传给上层。但调用napi的poll函数时在哪儿禁用对应设备的中断？
数据的发送： 两种方式： 1，dev_queue_xmit，会经过内核的流量控制子系统qdisc_run，最红调用hard_start_xmit 2,直接调用hard_start_xmit（）发送数据 net_tx_action()作用 1，释放完成队列的内存，2，调用qdisc_run对output队列上的数据
//初始化对应的函数指针 2627 netdev->open = e100_open; 2628 netdev->stop = e100_close; 2629 netdev->hard_start_xmit = e100_xmit_frame; 2630 netdev->set_multicast_list = e100_set_multicast_list; 2631 netdev->set_mac_address = e100_set_mac_address; 2632 netdev->change_mtu = e100_change_mtu; 2633 netdev->do_ioctl = e100_do_ioctl; 2634 SET_ETHTOOL_OPS(netdev, &e100_ethtool_ops); 2635 netdev->tx_timeout = e100_tx_timeout; 2636 netdev->watchdog_timeo = E100_WATCHDOG_PERIOD;
//napi支持，分配napi结构，初始化，并挂在netdevice的napi_list上 2642 nic = netdev_priv(netdev); 2643 netif_napi_add(netdev, &nic->napi, e100_poll, E100_NAPI_WEIGHT);
//nic的初始化 2702 init_timer(&nic->watchdog); 2703 nic->watchdog.function = e100_watchdog; 2704 nic->watchdog.data = (unsigned long)nic; 2705 init_timer(&nic->blink_timer); 2706 nic->blink_timer.function = e100_blink_led; 2707 nic->blink_timer.data = (unsigned long)nic; 2708 2709 INIT_WORK(&nic->tx_timeout_task, e100_tx_timeout_task);
//向内核注册 register_netdev(netdev)
当驱动激活的时候调用e100_open()函数 //初始化和分配rxs数据接收链表 2124 if((err = e100_rx_alloc_list(nic))) 2125 return err; //分配和初始化cbs命令链表，数据的发送通过这个 2126 if((err = e100_alloc_cbs(nic))) 2127 goto err_rx_clean_list; 2128 if((err = e100_hw_init(nic))) 2129 goto err_clean_cbs; 2130 e100_set_multicast_list(nic->netdev); 2131 e100_start_receiver(nic, NULL); 2132 mod_timer(&nic->watchdog, jiffies); 2133 if((err = request_irq(nic->pdev->irq, e100_intr, IRQF_SHARED, 2134 nic->netdev->name, nic->netdev))) 2135 goto err_no_irq; 2136 netif_wake_queue(nic->netdev); 2137 napi_enable(&nic->napi); 数据的发送 static int e100_xmit_frame(struct sk_buff *skb, struct net_device *netdev) 1637 { 1650 err = e100_exec_cb(nic, skb, e100_xmit_prepare); 1651 1652 switch(err) { 1653 case -ENOSPC: 1654 /* We queued the skb, but now we're out of space. */ 1655 DPRINTK(TX_ERR, DEBUG, "No space for CB\n"); 1656 netif_stop_queue(netdev); 1657 break; 1658 case -ENOMEM: 1659 /* This is a hard error - log it. */ 1660 DPRINTK(TX_ERR, DEBUG, "Out of Tx resources, returning skb\n"); 1661 netif_stop_queue(netdev); 1662 return 1; 1663 } 1664 1665 netdev->trans_start = jiffies; 1666 return 0; 1667 } 通过e100_exec_cb()发送数据，并对发送的错误进行处理，如果发送失败，并net_device的发送队列设置相应的标志，参见netif_stop_queue。 其中 e100_xmit_prepare对cb作一些设置，e100_exec_cb才发送数据. 1619 static void e100_xmit_prepare(struct nic *nic, struct cb *cb, 1620 struct sk_buff *skb) 1621 { 1622 cb->command = nic->tx_command; 1623 /* interrupt every 16 packets regardless of delay */ 1624 if((nic->cbs_avail & ~15) == nic->cbs_avail) 1625 cb->command |= cpu_to_le16(cb_i); 1626 cb->u.tcb.tbd_array = cb->dma_addr + offsetof(struct cb, u.tcb.tbd); 1627 cb->u.tcb.tcb_byte_count = 0; 1628 cb->u.tcb.threshold = nic->tx_threshold; 1629 cb->u.tcb.tbd_count = 1; 1630 cb->u.tcb.tbd.buf_addr = cpu_to_le32(pci_map_single(nic->pdev, 1631 skb->data, skb->len, PCI_DMA_TODEVICE)); 1632 /* check for mapping failure? */ 1633 cb->u.tcb.tbd.size = cpu_to_le16(skb->len); 1634 }
e100_exec_cb()则对nic上的cb链表进行处理 837 static int e100_exec_cb(struct nic *nic, struct sk_buff *skb, 838 void (*cb_prepare)(struct nic *, struct cb *, struct sk_buff *)) 839 { //使用一个空闲cb结构 851 cb = nic->cb_to_use; 852 nic->cb_to_use = cb->next; 853 nic->cbs_avail--; 854 cb->skb = skb; //调用e100_xmit_prepare对cb进行设置 859 cb_prepare(nic, cb, skb);
//对链表上的cb进行处理 867 while(nic->cb_to_send != nic->cb_to_use) { 868 if(unlikely(e100_exec_cmd(nic, nic->cuc_cmd, 869 nic->cb_to_send->dma_addr))) { 870 /* Ok, here's where things get sticky. It's 871 * possible that we can't schedule the command 872 * because the controller is too busy, so 873 * let's just queue the command and try again 874 * when another command is scheduled. */ 875 if(err == -ENOSPC) { 876 //request a reset 877 schedule_work(&nic->tx_timeout_task); 878 } 879 break; 880 } else { 881 nic->cuc_cmd = cuc_resume; 882 nic->cb_to_send = nic->cb_to_send->next; 883 } 884 }
e100_exec_cmd执行具体的操作，到这儿，哥就不懂了，有很多硬件相关的东西。有那些cmd?dma起了什么作用？
数据的接受 注册的中断函数 e100_intr，经典的中断处理例程 2032 static irqreturn_t e100_intr(int irq, void *dev_id) 2033 { 2034 struct net_device *netdev = dev_id; 2035 struct nic *nic = netdev_priv(netdev); //读取寄存器，判断中断是我们产生的还是其他设备产生的 2036 u8 stat_ack = ioread8(&nic->csr->scb.stat_ack); 2037 2038 DPRINTK(INTR, DEBUG, "stat_ack = 0x%02X\n", stat_ack); 2039 2040 if(stat_ack == stat_ack_not_ours || /* Not our interrupt */ 2041 stat_ack == stat_ack_not_present) /* Hardware is ejected */ 2042 return IRQ_NONE; 2043 //ack 我们的设备 2044 /* Ack interrupt(s) */ 2045 iowrite8(stat_ack, &nic->csr->scb.stat_ack); 2046 2047 /* We hit Receive No Resource (RNR); restart RU after cleaning */ 2048 if(stat_ack & stat_ack_rnr) 2049 nic->ru_running = RU_SUSPENDED; 2050 //把设备的poll放到softnet_data的poll_list上 2051 if(likely(netif_rx_schedule_prep(netdev, &nic->napi))) { 2052 e100_disable_irq(nic); 2053 __netif_rx_schedule(netdev, &nic->napi); 2054 } 2055 2056 return IRQ_HANDLED; 2057 } 数据的接收用的是napi方法，netif_rx_schedule()最终会调用napi_schedule把设备的poll方法挂在poll_list上。 2359 void __napi_schedule(struct napi_struct *n) 2360 { 2361 unsigned long flags; 2362 2363 local_irq_save(flags); 2364 list_add_tail(&n->poll_list, &__get_cpu_var(softnet_data).poll_list); 2365 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 2366 local_irq_restore(flags); 2367 } 设备的poll方法 2059 static int e100_poll(struct napi_struct *napi, int budget) 2060 { 2061 struct nic *nic = container_of(napi, struct nic, napi); 2062 struct net_device *netdev = nic->netdev; 2063 unsigned int work_done = 0; 2064 2065 e100_rx_clean(nic, &work_done, budget); 2066 e100_tx_clean(nic); 2067 2068 /* If budget not fully consumed, exit the polling mode */ 2069 if (work_done < budget) { 2070 netif_rx_complete(netdev, napi); 2071 e100_enable_irq(nic); 2072 } 2073 2074 return work_done; 2075 } 最终调用netif_recive_skb将数据包提交到上层。具体到硬件还是不太懂啊.：（
内核对数据的接收 有两种模式中断和NAPI。 先说核心的数据结构 996 /* 997 * Incoming packets are placed on per-cpu queues so that 998 * no locking is needed. 999 */ 1000 struct softnet_data 1001 { 1002 struct Qdisc *output_queue; 1003 struct sk_buff_head input_pkt_queue; 1004 struct list_head poll_list; 1005 struct sk_buff *completion_queue; 1006 1007 struct napi_struct backlog; 1008 #ifdef CONFIG_NET_DMA 1009 struct dma_chan *net_dma; 1010 #endif 1011 }; 正如注释所说，每个CPU都有一个，当涉及它的操作时，先关中断，所以不需要锁的保护 普通中断模式下，中断通过netif_rx把数据提交给input_pkt_queue,如果backlog没有添加到poll_list中，添加之。 而在NAPI下只是调用netif_rx_schedule判断是否需要把设备的poll方法添加到poll_list中. 添加后raise_softirq(NET_RX_ACTION),调用net_rx_action软中断进行后续的处理； net_rx_action遍历poll_list链表，调用其poll方法，对非NAPI设备来说就是通过process_backlog通用的poll方法把input_pkt_queue上的数据提交到上层，而对NAPI来说调用自身的poll方法。其中包含流量控制和对poll调用次数的限制。 process_backlog通过netif_receive_skb处理桥接，数据包监听，最后ip_rcv传给ip层 以e100_poll为例，其遍历rxs接收数据链表，对缓冲区的数据作unmap操作，然后调用netif_receive_skb传给上层。但调用napi的poll函数时在哪儿禁用对应设备的中断？
数据的发送： 两种方式： 1，dev_queue_xmit，会经过内核的流量控制子系统qdisc_run，最红调用hard_start_xmit 2,直接调用hard_start_xmit（）发送数据 net_tx_action()作用 1，释放完成队列的内存，2，调用qdisc_run对output队列上的数据
相关阅读 更多 +
