1. link状态分两端
+-------------+ | +--------+
| CPU | link1 +-------+ link2 | other |
| SGMII|<=====>| PHY |<==|==>| system |
| (MAC)| +-------+ | +--------+
+-------------+ |
当配置为SGMII模式时，它就是一个MAC；对应的，配置为1000base-x时，它是PHY；
对外面连的真正的PHY来说，有两个link状态：
- link1是指和MAC（PCS）层的link；
- link2是指和外部系统的link；好像这两个link都可以自协商
2. 关于phy
在open的时候connect phy， 在stop的时候disconnect phy
3. 代码中检测link状态
cvmx-helper.c：
cvmx_helper_link_info_t cvmx_helper_link_get(int ipd_port)
这个函数调用具体interface的link函数.
【注】这个函数用于没有phy的情况下
在ethernet-sgmii.c里面
cvm_oct_sgmii_open时
if (priv->phydev) 有phy时
int r = phy_read_status(priv->phydev);
cvm_oct_adjust_link(dev);
根据priv->phydev->link来改变link状态
else 没有phy时
link_info = cvmx_helper_link_get(priv->ipd_port);
priv->poll = cvm_oct_sgmii_poll; 起个pull来轮询
4. 相关命令
modprobe octeon-ethernet.ko
ifconfig eth0 192.168.0.8
ifconfig eth1 192.168.1.9
ifconfig eth0 up
ifconfig eth1 up
5. 设备注册及初始化
modprobe octeon-ethernet.ko
module_platform_driver(cvm_oct_driver); //相当于传统的module_init()和 module_exit(hello_exit)
module_driver(cvm_oct_driver, platform_driver_register, platform_driver_unregister)
platform_driver_register(struct platform_driver *drv=&cvm_oct_driver)
drv->driver.bus = &platform_bus_type;
//匹配设备时调用
drv->driver.probe = platform_drv_probe;
//设备移除时调用
drv->driver.remove = platform_drv_remove;
//关机时调用
drv->driver.shutdown = platform_drv_shutdown;
driver_register(&drv->driver);
//在fdt中找到匹配的设备后, 调用驱动提供的probe函数
cvm_oct_probe(struct platform_device *pdev)
//强制MDIO依赖
octeon_mdiobus_force_mod_depencency();
//fdt中的pip节
pip = pdev->dev.of_node;
cvm_oct_get_port_status(pip);
//70xx有5个interface, 每个interface又有n个port
//全局变量iface_ops
iface_ops[0] = &iface_ops_sgmii;
iface_ops[1] = &iface_ops_sgmii;
iface_ops[2] = &iface_ops_npi;
iface_ops[3] = &iface_ops_loop;
iface_ops[4] = &iface_ops_agl;
//对每个interface和port遍历
case CVMX_HELPER_INTERFACE_MODE_SGMII:
case CVMX_HELPER_INTERFACE_MODE_QSGMII:
//根据dts里面的关键字设置mac_phy_mode和1000x_mode
//只有SGMII和QSGMII有这两个选项
"cavium,sgmii-mac-phy-mode"
"cavium,sgmii-mac-1000x-mode"
//内核工作队列, 用于POLL
cvm_oct_poll_queue = create_singlethread_workqueue("octeon-ethernet");
cvm_oct_configure_common_hw();
//使能FPA
cvmx_fpa_enable();
//用FPA分配packet pool, 每个元素16个cache line; 一共8个pool
packet_pool = cvm_oct_alloc_fpa_pool(packet_pool, 16个cache line);
//1024个packet buffer, 默认值, 加载模块时可改num_packet_buffers
cvm_oct_mem_fill_fpa(packet_pool, num_packet_buffers);
//用FPA分配wqe pool
wqe_pool = cvm_oct_alloc_fpa_pool(-1, 一个cache line);
//用FPA分配output buffers pool
output_pool = cvm_oct_alloc_fpa_pool(-1, 8个cache line);
//初始化PIP, IPD, PKO. 硬件层初始化
cvmx_helper_initialize_packet_io_global();
cvmx_helper_interface_probe(int interface)
__cvmx_helper_init_interface(...)
cvmx_helper_ipd_and_packet_input_enable();
//对每个端口
__cvmx_helper_packet_hardware_enable(interface);
//详见下面第6条
iface_ops[interface]->enable(interface)
//对每个interface和每个port
//转换为IPD port, 将所有口加入到pow_receive_group, 默认15. 可选0-15
port = cvmx_helper_get_ipd_port(interface, index);
//使能IPD/PIP
cvmx_helper_ipd_and_packet_input_enable();
//对每个可用的口, 填写dev的各种信息, 注册dev, sgmii注册的是"eth0 eth1"
struct octeon_ethernet *priv;
struct net_device *dev;
dev = alloc_etherdev(sizeof(struct octeon_ethernet));
priv = netdev_priv(dev);
//填充priv结构
AGL口也使用SGMII的op
register_netdev(dev)
//将priv加入到全局链表cvm_oct_list
list_add_tail(&priv->list, &cvm_oct_list);
cvm_oct_add_ipd_port(priv);
queue_delayed_work(cvm_oct_poll_queue,
&priv->port_periodic_work, 5*HZ);
cvm_oct_rx_initialize()
queue_delayed_work(cvm_oct_poll_queue, &cvm_oct_rx_refill_work, HZ);
//也是个周期work, 1秒一次
void cvm_oct_rx_refill_worker(struct work_struct *work)
cvm_oct_rx_refill_pool(num_packet_buffers / 2);
queue_delayed_work(cvm_oct_poll_queue,
&cvm_oct_rx_refill_work, HZ);
...
...
...
...
6. 周期worker
在probe函数里, 挂了一个周期worker, INIT_DELAYED_WORK(&priv->port_periodic_work, cvm_oct_periodic_worker);, 5秒一次
void cvm_oct_periodic_worker(struct work_struct *work)
priv = container_of();
//这里的 priv->poll就是在接口open时注册的poll函数, 对SGMII来说, 就是cvm_oct_sgmii_poll
poll_fn = priv->poll;
poll_fn(priv->netdev);
//这里调用的是 struct net_device_stats *cvm_oct_common_get_stats(struct net_device *dev)
priv->netdev->netdev_ops->ndo_get_stats(priv->netdev);
queue_delayed_work(cvm_oct_poll_queue, &priv->port_periodic_work, 5*HZ);
7. sgmii初始化
cvm_oct_sgmii_init(struct net_device *dev)
struct octeon_ethernet *priv = netdev_priv(dev);
cvm_oct_common_init(dev);
从dts获取mac地址
如果没有mac地址, 则用random地址
SET_ETHTOOL_OPS(dev, &cvm_oct_ethtool_ops);
dev->netdev_ops->ndo_stop(dev);
//phy层检测到link change, 就调用这个函数
priv->link_change = cvm_oct_sgmii_link_change;
8. 当执行ifconfig eth0 up时
int cvm_oct_sgmii_open(struct net_device *dev)
cvm_oct_phy_setup_device(dev);
//从dts里面找phy-handle
phy_node = of_parse_phandle(priv->of_node, "phy-handle", 0)
//SGMII不走下面, rgmii走下面, cvm_oct_adjust_link是个回调函数
priv->phydev = of_phy_connect(dev, phy_node, cvm_oct_adjust_link, 0, iface)
phy_start_aneg(priv->phydev);
//有phy走if分支
if (priv->phydev)
r = phy_read_status(priv->phydev)
cvm_oct_adjust_link(dev);
//没有phy走else分支
else
link_info = cvmx_helper_link_get(priv->ipd_port)
//这个函数获取速率时有点问题, 需要定位
__cvmx_helper_sgmii_link_get()
根据板类型直接返回up(FPXTB)
priv->poll = cvm_oct_sgmii_poll;
//其实这个函数是周期性调的.
cvm_oct_sgmii_poll(dev);
link_info = cvmx_helper_link_get(priv->ipd_port)
//和以前比, 一样就直接返回了. 第一次不一样, 往下走
cvmx_helper_link_autoconf(priv->ipd_port)
link_info = cvmx_helper_link_get(ipd_port);
//更内部的link状态记录在cvmx_interfaces[interface].cvif_ipd_port_link_info
//如果上面获取的link_info和内部的cvif_ipd_port_link_info一样, 就直接返回.
//第一次肯定不一样, 往下走
cvmx_helper_link_set(ipd_port, link_info);
__cvmx_helper_sgmii_link_set()
//要禁止自协商, 在这里改. PCS层reset+配置
__cvmx_helper_sgmii_hardware_init_link(interface, index);
//主要是根据link_info配置GMX层寄存器, 速度, 全双工.
__cvmx_helper_sgmii_hardware_init_link_speed(interface, index, link_info)
9. 在加载模块时, 会进行硬件初始化
__cvmx_helper_sgmii_enable(int interface)
__cvmx_helper_sgmii_hardware_init(interface, num_ports);
__cvmx_helper_setup_gmx(interface, num_ports);
__cvmx_helper_sgmii_hardware_init_one_time(interface, index);
//初始化link状态
__cvmx_helper_sgmii_link_set(ipd_port, __cvmx_helper_sgmii_link_get(ipd_port))