Building TCP/IP packets

1. Building TCP/IP packets A look at the computation-steps which need to be performed for utilizing the TCP/IP protocol

2. NIC’s ‘offloading’ capabilities • An advanced feature of Intel’s most recent PRO1000 gigabit ethernet controllers is an ability to perform many of the calculations associated with the processing of network packets which otherwise would have to be programmed by device-driver authors and then executed at runtime by a CPU that has much other vital work to attend to

3. How much work? • To gain a concrete idea of the work that is required to set up a conventional network packet for transmission, and then to take it apart when it’s received, we have created a character-mode Linux driver that shows us all of the essential calculation-steps, as it does not rely on the usual networking software included within Linux’s kernel

4. TCP/IP over Ethernet Here is a diagram, from Intel’s Software Developer’s Manual, of the widely deployed TCP/IP Ethernet packet format

5. UnixWare: online tutorial recommended background reading “The TCP/IP Protocol Stack” © 2002 Caldera International, Inc. All rights reserved. UnixWare 7 Release 7.1.3 - 30 October 2002 <http://uw713doc.sco.com/en/NET_tcpip/tcpN.tcpip_stack.html>

6. TCP/IP Protocol Stack by Caldera, Incorporated

7. Encapsulation and Decapsulation of Data within a network stack by Caldera, Incorporated

8. Network stack support for TCP/IP by Caldera, Incorporated

9. Internet Protocol Header

10. IP Header (Little Endian layout)

11. Transport Control Protocol Header

12. TCP Header (Little-Endian layout)

13. TCP Pseudo Header layout

14. Our ‘nictcp.c’ demo • This device-driver’s ‘write()’ function will show the sequence of steps needed for building a packet with the TCP/IP format: • Setup the packet’s data ‘payload’ • Prepend a TCP packet-header • Then prepend a TCP ‘pseudo’ packet-header • Compute and insert the TCP checksum • Replace the pseudo-header with an IP header • Compute and insert the IP Header Checksum • Finally prepend the Ethernet Header

15. Two checksum calculations The TCP Checksum gets inserted here The TCP Segment TCP Pseudo-Header TCP Header packet-data The TCP Checksum is computed over this array of words The IP Header Checksum gets inserted here Frame Header IP Header TCP Header packet-data The IP Header Checksum is computed over this array of words

16. Sample checksum algorithm // to compute and insert the TCP Checksum unsigned char *cp = phys_to_virt( txring[ txtail ].base_addr ); unsigned short *wp = (unsigned short *)( cp + 22 ); unsigned int nbytes = 12 + 20 + len; unsigned int nwords = (nbytes / 2) + (nbytes % 2); unsigned int cksum = 0; if ( len & 1 ) cp[ 14 + 20 + 20 + len ] = 0; // padding for (int i = 0; i < nwords; i++) cksum += htons( wp[ i ] ); cksum += (cksum >> 16); // end-around-carries cksum & 0xFFFF; // mask for 16-bits cksum ^= 0xFFFF; // flip the low 16-bits *(unsigned short*)(cp + 50) = htons( cksum );

17. In-class demonstration • We can use our ‘pktsplit.c’ demo-module to confirm the correctness of our TCP/IP packet-format as far as the NIC hardware is concerned – it reports (in its descriptor ‘status’ and ‘errors’ fields) that both of our checksums were computed by the NIC on reception -- and that neither was in error

18. In-class exercise • Can you adapt these ideas in our ‘nictcp.c’ driver to produce the analogous ‘nicudp.c’ demonstration module? • A similar ‘pseudo’ UDP Header is used to calculate the UDP Checksum value (see the Intel Software Developer’s Manual for the layout and size of the UDP Header)