✎ soc:nikhilcrao - Etherboot/gPXE Wiki

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== Nikhil Rao, Implementing IPv6 support in gPXE ======

===== About the project =====
My project is titled Implementing IPv6 support in gPXE. Here are some snippets from [[soc:nikhil:proposal|my proposal]].

==== Synopsis ====
gPXE uses uIP to implement its TCP/IP stack. uIP is an implementation of the TCP/IP stack which uses a fraction of the resources used by a full fledged TCP/IP implementation.

The project is to implement support for IPv6. uIP (in its current avatar) tightly integrates the TCP and IP layers in order to save code size. As a result, it is not easy to replace IPv4 with any other network layer protocol in uIP. This project aims to re-design the TCP/IP stack of gPXE in order to facilitate IPv6 implementation.

==== Deliverables ====
The final deliverables for the project are:
  - A clean, well-defined interface between the transport-layer and network-layer. The interface would be generic enough to allow any transport layer protocol to interact with any network layer protocol
  - Re-design and implementation of the TCP/IP stack using the proposed interface (which fits within the gPXE API)
  - Minimal support for IPv6

The stretch goals for the project are:
  - Extension of the gPXE API to include UDP
  - An implementation of UDP/IP (within the extended gPXE API)
  - Support for additional features in IPv6

==== Plan of action ====

Main goals:
  - <del>Investigate the working of uIP</del>
  - <del>Define bare necessary requirements of transport layer and network layer</del>
  - <del>Define the interface between these layers based on the requirements</del>
  - Implement the TCP, IP modules using the interface
  - Test implementation; Re-implement if necessary
  - Increase requirements if necessary and go back to step 2
  - Investigate minimum requirements to support IPv6
  - Extend interface/requirements if necessary and go back to step 2
  - Implement IPv6
  - If time permits, implement stretch goals


Stretch goals:
  - Investigate UDP implementation in Etherboot-5.4 and earlier
  - Extend interface/requirements if necessary and perform steps 2 - 5
  - Implement UDP support
  - Investigate various features that can be added to IPv6
  - Perform steps 7 - 8

===== Current status =====

I have tried to update my blog ss frequently as possible with my thoughts. I have listed below my understanding of the problem.

==== Working of the network infrastructure ====

=== Physical Layer ===

Let us assume we are working with a RTL8139 driver. Further let us assume that we are using the Ethernet link layer protocol and the uIP stack to implement TCP/IP in gpxe. This short note will describe how data is received from the driver, queued and processed by gpxe's network stack.

I am not too sure about the architecture of the RTL8139 driver. According to my understanding (and a quick perusal of src/drivers/net/rtl8139.c), RTL8139 maintains a buffer for receiving packets. The structure of this buffer is:

  struct rtl8139_rx {
    void *ring;
    unsigned int offset;
  };

The RTL8139 NIC structure contains one such buffer for receiving packets and a similar buffer for transmitting packets:

  struct rtl8139_nic {
    struct threewire eeprom;
    unsigned short ioaddr;
    struct rtl8139_tx tx;
    struct rtl8139_rx rx;
  };

There are a bunch of functions to perform various tasks, such as opening the NIC, reading the MAC address, resetting the NIC, closing the NIC, etc. static void rtl_poll(struct net_device *netdev) is used to poll RTL8139 to check for received packets. This function takes a network device as an argument. The private data of the network device stores the rtl8139_nic structure. If data is available, it allocates a packet buffer of the appropriate size and copies the data from the driver into the packet buffer. It then calls void netdev_rx(struct net_device *netdev, struct pk_buff *pkb) passing the network device and packet buffer as arguments. The function netdev_rx() performs a very simple task. It fills up ll_protocol of the packet buffer with information from netdev and then adds the packet buffer to the rx_queue. The packet is picked up for processing by the link layer protocol (IPv4) when int net_rx_process() is called.

=== Link Layer ===

The packet buffer is an interesting concept detailed in src/include/gpxe/pkbuff.h and src/net/pkbuff.c. The structure used to represent a network packet within gpxe is:

  struct pk_buff {
    void *head;
    void *data;
    void *tail;
    void *end;
    struct list_head list;
    struct net_protocol *net_protocol;
    struct ll_protocol *ll_protocol;
  };


The first four pointers are used to demarcate the buffer and data boudaries. The list structure maintains a list of all devices/entities this buffer is a member of (such as rx_queue, etc.). The net_protocol and ll_protocol pointers point to the network and link layer protocols respectively.

Now, the interesting part is in the buffer allocation function, struct pk_buff * alloc_pkb(size_t len), given in src/net/pkbuff.c. It takes the requested length of the buffer as an input argument. It is usually called from rtl_poll() while checking for received data. The packet buffer structure comes immediately after the data. The allocation can be represented abstractly as:

{{:soc:pkbuff.jpg|:soc:pkbuff.jpg}}

A bunch of functions are provided in ''src/include/gpxe/pkbuff.h'' and ''src/net/pkbuff.c'' for appending, prepending, etc. data.

A network device is represented using the following structure (src/include/gpxe/netdevice.h):

  struct net_device {
    int (* transmit) (struct net_device *netdev, struct pk_buff *pkb);
    void (* poll) (struct net_device *netdev);
    struct ll_protocol *ll_protocol;
    uint8_t ll_addr[MAX_LL_ADDR_LEN];
    void *priv;
  };

Every network device has to implement two functions - transmit() and poll() which send data and poll for new data respectively. A network device is also associated with a link layer protocol (which it implements) and a link layer address. If the link layer protocol is Ethernet, then the link layer address it the MAC address of the network interface. The link layer protocol is represented as (src/include/gpxe/netdevice.h):

  struct ll_protocol {
    const char *name;
    int (* route) (struct net_device *netdev, const struct net_header *nethdr, struct ll_header *llhdr);
    void (* fill_llh) (const struct ll_header *llhdr, struct pk_buff *pkb);
    void (* parse_llh) (const struct pk_buff *pkb, struct ll_header *llhdr);
    const char* (*ntoa) (const void *ll_addr);
    uint16_t ll_proto;
    uint8_t ll_addr_len;
    uint8_t ll_header_len;
  };

Every link layer protocol has to implement route(), fill_llh(), parse_llh and ntoa(). The function route() performs link layer routing. It constructs a generic link layer header llhdr from a generic network layer header nethdr. The function fill_llh() is used to fill the media-specific link layer header. Similarly parse_llh() parses the link layer header in the packet buffer and fills in the generic link layer header llhdr. ntoa() is used to represent a link layer address in a human readable format.

The ethernet protocol implements these functions in src/net/ethernet.c as eth_route(), eth_fill_llh(), eth_parse_llh() and eth_ntoa() respectively.

Currently, uIP supports only a single instance of a network device. This single instance is called static_single_netdev within the code (refer src/net/netdev.c). The functions alloc_netdev() and register_netdev() should still be used to allocate and register a network device (although in the current implementation the code will be optimized out). There is a received packet queue, rx_queue, which maintains a list of received packet buffers. In the current setup, netdev->poll() is set to point to rtl_poll() when the RTL network device is probed in rtl_probe() (refer src/drivers/net/rtl8139.c).

A single step network operation is peformed by calling static void net_step(struct process *process). This function polls all the network devices for new packets using int net_poll(). This function polls for packets on all network devices by calling netdev->poll() and returns true if there are packets present in the receive queue (rx_queue in our case). net_step() handles at most one received packet at a time. It dequeues the received packet using struct pk_buff * net_rx_dequeue() and then processes the received packet using int net_rx_process(struct pk_buff *pkb). After this processing is complete, it schedules itself using the schedule() function.

The process function, int net_rx_process(struct pk_buff *pkb), processes a received packet at the link layer. Note that the link layer protocol is specified in the packet buffer in the field ll_protocol. This function fills up a generic link layer header llhdr by parsing the media specific components of the link layer header in the packet. The link layer header is represented as (refer src/include/gpxe/netdevice.h):

  struct ll_header {
    struct ll_protocol *ll_protocol;
    int flags;
    uint8_t dest_ll_addr[MAX_LL_ADDR_LEN];
    uint8_t source_ll_addr[MAX_LL_ADDR_LEN];
    uint16_t net_proto;
  };

A generic link layer header consists of a pointer to the link layer protocol (in this case, to the ethernet link layer protocol), a field for flags (which is a bitwise OR of zero or more PKT_FL_XXX values, a destination and source link layer address (in the case of ethernet, a 48 bit address, where MAX_LL_ADDR_LEN = 6) and a 16 bit identification of the network protocol in the IP datagram.

The files src/include/gpxe/ethernet.h and src/net/ethernet.c contain the specifications for the ethernet protocol along with definitions for the various methods the protocol has to implement. The ethernet header is represented as (refer src/include/gpxe/if_ether.h):

  struct ethhdr {
    uint8_t h_dest[ETH_ALEN];
    uint8_t h_source[ETH_ALEN];
    uint16_t h_protocol;
  };

The ethernet header is mapped on to the link layer header of the received packet and the corresponding fields are copied in to the generic link layer header llhdr. net_rx_process() then identifies the network layer protocol using the struct net_protocol * find_net_protocol(int net_proto) function and passing llhdr.net_proto as an argument to it. It sets the network protocol field net_protocol in the packet buffer to the protocol returned by the find_net_protocol(). It then strips off the link layer header and hands the packet buffer to the network layer (IP) to process by calling int net_protocol->rx_process(struct pk_buff *pkb).

=== Network Layer ===

A network protocol is represented as (src/include/gpxe/netdevice.h):

  struct net_protocol {
    const char *name;
    int (* route) (const struct pk_buff *pkb, struct net_header *nethdr);
    int (* rx_process) (struct pk_buff *pkb);
    const char* (*ntoa) (const void *net_addr);
    uint16_t net_proto;
    uint8_t net_addr_len;
  };

Every network protocol has to implement the functions route(), rx_process() and ntoa(). The function route() performs network layer routing. It fills in the network header nethdr with enough information to allow the link layer to route the packet. The function rx_process() processes a received packet and ntoa() represents the network address in a human readable format.

The IPv4 protocol implements these functions in src/net/ipv4.c as ipv4_route(), ipv4_rx() and ipv4_ntoa() respectively.

In the current setup, the packet is handed over to uIP to process at this step. When net_protocol->rx_process() is called, the caller passes the packet buffer as an argument. The uIP stack is set up and the packet is copied into uip_buf as specified by uIP. The function uip_input() is then called and the packet is processed.

When uip_input() returns, it could have some data in uip_buf (which needs to be sent out). I will cover this in the next section on sending data.

== uIP TCP/IP stack ==

The uIP module is defined in src/net/uip/uip.h, src/net/uip/uipopt.h, src/net/uip/uip_arch.h, src/net/uip/uip.c and src/net/uip/uip_arch.c. Internally, uip_input() calls the TCP/IP state machine (implemented in uip_process()) with UIP_DATA passed into it as an argument. This indicates that a packet has been received and needs to be processed.

The uip_process() function is split into two parts - one that handles periodic firings of uip_process() and another that handles input processing. The second part is invoked in this case. The IPv4 header is processed in the following steps:

  - Check validity of IPv4 header. As uIP does not process options, it expects the header length to be of length 5 measured as 32 bit words.
  - Check the size of the packet to ensure uip_len and the length specified in the IP header are the same.
  - Check if the received packet is a fragmented packet. If so, then call uip_reass() which reassembles the fragments [and exit].
  - Do some ICMP processing if we are configured to use ping IP address configuration and our IP address is 0.0.0.0
  - Check if the packet is destined for our IP address. If not, then drop the packet.
  - Compute and check the checksum of the IP header.
  - Check the transport layer protocol and invoke the appropirate module.

Currently the uIP stack supports only TCP, UDP and ICMP.

**ICMP Processing**

uIP is set up to handle only ICMP_ECHO (and, if configured, ICMP_PINGADDRCONF) processing. In ICMP_ECHO processing, the type of the ICMP message is changed to ICMP_ECHO_REPLY, the checksum is calculated and addresses are swapped. The ICMP packet is placed on the buffer. Note that the length of the buffer, uip_len, is not changed since the outgoing packet is the same size as the incoming packet. The function returns and this packet is sent (refer the next section on sending data).

=== Transport Layer ===

**UDP Processing**

The UDP processing of uIP does not do anything to the UDP/IP headers. It sends the information back to the UDP application which does all the hard work. The UDP state machine checks the checksum of the UDP packet if it is configured to do so. It then proceeds to check which UDP connection the packet should go to. If it finds a connection, it strips the UDP header, sets the appropriate flags and sends it to the application via UIP_UDP_APPCALL().

If the application wants to send data it places the data in the app_data buffer and sets uip_slen to the length of the data. When uIP returns, it checks if uip_slen is non-zero which indicates that the application wants to send some data. It then proceeds to filling in the transport and network layer headers.

**TCP Processing**

TCP processing proceeds in the following steps:

  - Compute and check the TCP checksum
  - Demux this TCP segment between the TCP connections; depending on the type of packet, process accordingly
    - Check all active connections that are expecting a SYN,ACK packet after sending a SYN packet; if found, go to 4.
      - Check if it is the SYN flag is set; if so, then it is an old duplicate - send a RST and exit.
      - Check all listen connections to see if the destination ports match; if nothing is found send a RST and exit.
  - If incoming packet is intended for a listening port (2.c)
      - Search for a empty connection
      -  Fill in all necessary fields
      - Change connection state to SYN_RCVD
      - Check TCP MSS option if available and use it to set MSS
      - Send SYN,ACK packet
  - If incoming packer is intended for an active connection (2.a)
      - Check TCP RST and reset connection if set
      - Calculate the length of the data send to us
      - Check if the SEQ_NUM of the incoming packet is what we are expecting next; if not, send an ACK with the correct numbers in it.
      - Check if the incoming segment ACKs any outstanding data; if so, update SEQ_NUM, reset the length of the outstanding data, calculate RTT estimations and reset the timer
      - Switch based on the TCP state of the connection:
        * CASE SYN_RCVD:
	  - If connection state is ACKDATA, change TCP state to ESTABLISHED; if not, drop the packet.
	  - Change connection state to CONNECTED
	  - If there is any data in the packet, put it in the buffer and set the connection state to NEWDATA
	  - Call the application
	* CASE SYN_SENT:
	  - if the SYN and ACK flags are set and the connection is in the ACKDATA state, proceed
	  - Parse the MSS option if present
	  - Set the TCP state to ESTABLISHED
	  - Set the connection state to CONNECTED | NEWDATA
	  - Call the application
	* CASE ESTABLISHED:
	  - If the packet is a FIN and there is no outstanding data, then close the connection and inform application
	  - Check URG flag to process urgent data
	  - If uip_len > 0, we have new data; set the connection state to NEWDATA and update the SEQ_NUM we acknowledge
	  - If the application has stopped dataflow using uip_stop() do not accept any new data
	  - Set the MSS
	  - Call the application
	* CASE LAST_ACK:
	  - If ACK is received, then close the connection and call the application
	* CASE FIN_WAIT_1:
	  - ...
	* CASE FIN_WAIT_2:
	  - ...
	* CASE TIME_WAIT:
	* CASE CLOSING:

=== Application Layer ===

The application layer is invoked through the gPXE TCP API.

** Todo: Add sending data **

=== So, what's wrong? ===

The uIP TCP/IP stack, as mentioned in my proposal, is very tightly packed. The entire processing takes place in one function uip_process(). As mentioned earlier, this is split into two parts - one for processing new data and another to periodically check timeouts and see if data is to be sent. uip_process() heavily relies on ''goto'' statements and as a result, adding support for a new protocol is a difficult.

Internally, we use packet buffers to hold packets and information about it. When the uIP stack is invoked, data is transfered from the packet buffer into the uip_buffer (which is a statically allocated space for the data in uIP). This is inefficient usage of memory.



==== Redesigning the IP layer ====

=== IPv4 ===

==== Redesigning the transport layer ====

=== UDP ===

=== TCP ===

===== The road ahead =====

==== Ideas and concepts ====

==== To do list ====

===== Resources =====

  * [[http://implementingipv4.blogspot.com|My blog]]
  * [[http://www.sics.se/~adam/uip/|uIP Project Page]]
  * [[http://playground.sun.com/pub/ipng/html/ipng-main.html|IPv6 resources on sun.com]]
Trace: • dnsmasq • tg3 • freedos • week10 • 2009 • gentoo_iscsi • week6 • debian_lenny_iscsi • opensolaris_iscsi_install • stefanha • winnt_sanbootconf • week12 • winnt_aoe • week3 • dslbko • win2k8_iscsi_install • ubuntu_iscsi • kylek