2015-05-22

how tcpdown direction filter work

tcpudmp对 direction的支持。

内核代码有一个关键数据结构：skb的pkt_type字段。
在收发路径这个域被赋值为PACKET_OUTGOING或者其他。
这个值被传递到往用户空间，libpcap根据它判断报文的方向是否是期望的。

pkt_type的可能取值

24 #define PACKET_HOST             0               /* To us                */
25 #define PACKET_BROADCAST        1               /* To all               */
26 #define PACKET_MULTICAST        2               /* To group             */
27 #define PACKET_OTHERHOST        3               /* To someone else      */
28 #define PACKET_OUTGOING         4               /* Outgoing of any type */

2015-05-21

virtnet

vhost net study

vhost net 的目的是为了避免在host kerne上做一次qemu的调度，提升性能。
xmit: 让vm的数据报在 host的内核就把报文发送出去。
rcv：

核心数据结构

`vhost_poll`是vhost里最关键的一个数据结构。

27 /* Pol> > file (eventfd or socket) */
28 /* Note: there's nothing vhost specific about this structure. */
29 struct vhost_poll {
3> > > >   poll_tabl> > > > > > > >  table;
3> > > >   wait_queue_head_> > > >  *wqh;
3> > > >   wait_queue_> > > > > >    wait;
3> > > >   struct vhost_wor> > > >   work;
3> > > >   unsigned lon> > > > > >   mask;
3> > > >   struct vhost_de> > > >   *dev;
36 };

table:每次负责把wait域放倒wqh里。vhost_net_open将它的执行函数vhost_poll_func
wqh：它的wqh被初始化指向一个eventfd的ctx，
wait：每次把wait放倒这个wqh链表里，当guest vm的发送报文时，wait被摘下，
并执行其对应的func,vhost_net_open将该func被初始化为vhost_poll_wakeup。
vhost_poll_wakeup负责将work放入对应vhost_dev下的work_list链表中。
work: 每个vhost_dev有一个thread，负责从work_list链表里的摘除work节点，
并执行work节点对应的fn. fn是真正干活的的函数。
对于rx vhost_virqueue, vhost_net_open将该fn初始化为handle_rx_kick
对于tx vhost_virqueue, vhost_net_open将该fn初始化为handle_tx_kick
对于rx vhost_virqueue, vhost_net_open将该fn初始化为
handle_rx_kick.
mask:是需要监听的eventfd的事件集合
dev: 该vhost_poll对应的vhost_dev;

2015-05-21

route

ipvlan study

L2 mode

xmit packet

xmit a normal pkt to other phy machine

==> ipvlan_start_xmit
==> ==> ipvlan_xmit_mode_l2
==> ==> ==> skb->dev = ipvlan->phy_dev;
==> ==> ==> return dev_queue_xmit(skb);

xmit a normal pkt to other namespace

==> ipvlan_start_xmit
==> ==> ipvlan_xmit_mode_l2
==> ==> ==> if (ether_addr_equal(eth->h_dest, eth->h_source))
==> ==> ==> addr = ipvlan_addr_lookup(ipvlan->port, lyr3h, addr_type, true);
==> ==> ==> ipvlan_rcv_frame(addr, skb, true);
==> ==> ==> ==> skb->dev = dev; <== dst namespace dev
==> ==> ==> ==> dev_forward_skb(ipvlan->dev, skb)
==> ==> ==> ==> ==> return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
==> ==> ==> ==> ==> ==> enqueue_to_backlog(skb, get_cpu(), &qtail);

xmit a mutlicast pkt

==> ipvlan_start_xmit
==> ==> ipvlan_xmit_mode_l2
==> ==> ==> else if (is_multicast_ether_addr(eth->h_dest))
==> ==> ==> ipvlan_multicast_frame(ipvlan->port, skb, ipvlan, true);
==> ==> ==> ==> list_for_each_entry(ipvlan, &port->ipvlans, pnode)
==> ==> ==> ==> ==> dev_forward_skb or netif_rx(nskb);

recv packet

All the packet are get by the rx_handler, ipvlan_handle_frame.

unicast packet

lookup the dest ipvlan port(net_device)
by the dst IPv4/6 address, and send to it.

==> ipvlan_handle_frame
==> ==> ipvlan_handle_mode_l2
==> ==> ==> ipvlan_addr_lookup(port, lyr3h, addr_type, true);
==>  ==> ==> ==> skb->dev = dev;
==> ==> ==> ==> dev_forward_skb or ret = RX_HANDLER_ANOTHER;

multicast packet.

==> ipvlan_handle_frame
==> ==> ipvlan_handle_mode_l2
==> ==> ==> if (is_multicast_ether_addr(eth->h_dest)) {
==> ==> ==> ipvlan_addr_lookup(port, lyr3h, addr_type, true);
==> ==> ==> ==> if (ipvlan_external_frame(skb, port))
==> ==>  ==> ==> ==> ipvlan_multicast_frame(port, skb, NULL, false);

l3 mode

1	ipvlan_start_xmit

308 static const struct net_device_ops ipvlan_netdev_ops = {
309         .ndo_init               = ipvlan_init,
310         .ndo_uninit             = ipvlan_uninit,
311         .ndo_open               = ipvlan_open,
312         .ndo_stop               = ipvlan_stop,
313         .ndo_start_xmit         = ipvlan_start_xmit,
314         .ndo_fix_features       = ipvlan_fix_features,
315         .ndo_change_rx_flags    = ipvlan_change_rx_flags,
316         .ndo_set_rx_mode        = ipvlan_set_multicast_mac_filter,
317         .ndo_get_stats64        = ipvlan_get_stats64,
318         .ndo_vlan_rx_add_vid    = ipvlan_vlan_rx_add_vid,
319         .ndo_vlan_rx_kill_vid   = ipvlan_vlan_rx_kill_vid,
320 };

495 int ipvlan_queue_xmit(struct sk_buff *skb, struct net_device *dev)
496 {
497         struct ipvl_dev *ipvlan = netdev_priv(dev);
498         struct ipvl_port *port = ipvlan_port_get_rcu(ipvlan->phy_dev);
499
500         if (!port)
501                 goto out;
502
503         if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
504                 goto out;
505
506         switch(port->mode) {
507         case IPVLAN_MODE_L2:
508                 return ipvlan_xmit_mode_l2(skb, dev);
509         case IPVLAN_MODE_L3:
510                 return ipvlan_xmit_mode_l3(skb, dev);
511         }
512
513         /* Should not reach here */
514         WARN_ONCE(true, "ipvlan_queue_xmit() called for mode = [%hx]\n",
515                           port->mode);
516 out:
517         kfree_skb(skb);
518         return NET_XMIT_DROP;
519 }

457 static int ipvlan_xmit_mode_l2(struct sk_buff *skb, struct net_device *dev)
458 {
459         const struct ipvl_dev *ipvlan = netdev_priv(dev);
460         struct ethhdr *eth = eth_hdr(skb);
461         struct ipvl_addr *addr;
462         void *lyr3h;
463         int addr_type;
464
465         if (ether_addr_equal(eth->h_dest, eth->h_source)) {
466                 lyr3h = ipvlan_get_L3_hdr(skb, &addr_type);
467                 if (lyr3h) {
468                         addr = ipvlan_addr_lookup(ipvlan->port, lyr3h, addr_type, true);
469                         if (addr)
470                                 return ipvlan_rcv_frame(addr, skb, true);
471                 }
472                 skb = skb_share_check(skb, GFP_ATOMIC);
473                 if (!skb)
474                         return NET_XMIT_DROP;
475
476                 /* Packet definitely does not belong to any of the
477                  * virtual devices, but the dest is local. So forward
478                  * the skb for the main-dev. At the RX side we just return
479                  * RX_PASS for it to be processed further on the stack.
480                  */
481                 return dev_forward_skb(ipvlan->phy_dev, skb);
482
483         } else if (is_multicast_ether_addr(eth->h_dest)) {
484                 u8 ip_summed = skb->ip_summed;
485
486                 skb->ip_summed = CHECKSUM_UNNECESSARY;
487                 ipvlan_multicast_frame(ipvlan->port, skb, ipvlan, true);
488                 skb->ip_summed = ip_summed;
489         }
490
491         skb->dev = ipvlan->phy_dev;
492         return dev_queue_xmit(skb);
493 }

2014-10-21

debug

ftrace study

test case

We found ixgbe rx softirq poll function ixgbe_poll was called even without pkt coming.
how to prove it and who call it?

analysis

By browse source ixgbe_poll is called, it should be done by napi schedule.
If this is true, __napi_schedule should be called.

2014-09-17

netcore

netdevice watchdog cause tx queue schedule

test case

For ixgbe nic, we want to assign a tx hardware qeueue to each cpu,
and the tx softirq should use the corresponding hardware queue.

each packet will select a softqueue in dev_queue_xmit,
we rewrite ixgbe driver ndo_select_queue(ixgbe_select_queue),
which will return current cpu index(based 0) when packet select queue.
thus for each cpu use its own tx queue.

but, we found some packet had unmatched queue index when send
on specific cpu.

for example, a packet’s queue index is 5 but is sent by cpu3,
thus, cpu3 will operate tx hw queue5, which should only be done by cpu5.

2014-09-16

how does ixgbe use queue index

data structure

7913 static const struct net_device_ops ixgbe_netdev_ops = {
...
7916         .ndo_start_xmit         = ixgbe_xmit_frame,
7917         .ndo_select_queue       = ixgbe_select_queue,
...

receive skb: record receive queue index

record queue_index +1, 0 is used as NOT record.

####call trace

> ixgbe_poll
> > ixgbe_clean_rx_irq
> > > ixgbe_process_skb_fields
> > > > skb_record_rx_queue

2014-09-14

netcore

tcpdump work with bonding interface

test case

####On redhat5, Why tcpdump could not work on bonding work.
OS: redhat 5.
There are two 82599 interfaces eth0 and eth1.
These two interfaces are used as slave of bond0,
eth1 is backup of eth0.

We ping the default gateway on test machine.
ping work OK, and tcpdump on bond0 show the icmp request and icmp require packets.
while on eth0 only icmp request, and eth1 has no any packet.

2014-07-31

socket

how to select source ip for a tcp socket

This question should be divided into two parts.

Part 1: adding a route

Add a direct(link) route

Add a route with gateway

create a socket

for example ‘connect’ system call

2014-07-17

IRQs

register irq handler

call trace

以handle_level_irq为例说明.

===> handle_level_irq
==> ==> handle_irq_event
==> ==> ==> handle_irq_event_percpu
==> ==> ==> ==>action->handler

where handler is registered

127 static inline int __must_check
128 request_irq(unsigned int irq, irq_handler_t handler, unsigned long flags,
129             const char *name, void *dev)
130 {
131         return request_threaded_irq(irq, handler, NULL, flags, name, dev);
132 }

2014-07-17

IRQs

irq framework

中断处理过程：

硬件中断到中断控制器

reg value–>irq(int) —> struct irq_desc

==> 中断时的有一个寄存器会保存中断源的vector值.
==> ==> `arch/x86/kernel/entry_64.S`调用函数`do_IRQ`.
==> ==> ==> `do_IRQ`依据`vector_irq`和vector值, 找到对应的中断号,并调用`handle_irq`.
==> ==> ==> ==> `handle_irq`通过函数irq_to_descdesc,可将中断号转化为`struct irq_desc`.
==> ==> ==> ==> generic_handle_irq_desc(irq, desc);
==> ==> ==> ==> ==> `generic_handle_irq_desc`调用 desc->handle_irq(irq, desc);

注:这里的handle_irq不是真正的中断处理函数,而是几大类中断控制器处理函数.
如82599, msi等.
具体分析见:irq study1

中断控制器到具体的中断处理函数

==> handle_level_irq
==> ==> irqreturn_t handle_irq_event(struct irq_desc *desc)
==> ==> ==> struct irqaction *action = desc->action
==> ==> ==> ret = handle_irq_event_percpu(desc, action);
==> ==> ==> ==> action->handler(irq, action->dev_id);

这里的action->handler才是我们使用request_irq注册的中断处理函数.
具体分析见:
具体分析见:irq study2

tcpudmp对 direction的支持。

pkt_type的可能取值

核心数据结构

vhost_poll是vhost里最关键的一个数据结构。

L2 mode

xmit packet

xmit a normal pkt to other phy machine

xmit a normal pkt to other namespace

xmit a mutlicast pkt

recv packet

unicast packet

multicast packet.

l3 mode

test case

analysis

test case

data structure

receive skb: record receive queue index

test case

Part 1: adding a route

Add a direct(link) route

Add a route with gateway

create a socket

for example ‘connect’ system call

call trace

where handler is registered

中断处理过程：

硬件中断到中断控制器

中断控制器到具体的中断处理函数

`vhost_poll`是vhost里最关键的一个数据结构。