Outline: Linux Architecture

• Linux overview
• overall architecture
• networking
• comparison to X-kernel, Xinu
• organization
• packet send, receive walkthrough

Linux Overview

• ancestry:
  • Unix (Bell Labs): system calls, utilities
  • Unix (Berkeley): networking, utilities
  • Unix (Posix): standardization, modules?
  • X windows: window system
  • GNU: open source, gcc, utilities
  • Minix: open source, initial impetus
• contributions:
  • free Unix -- unlocking the monopoly?
  • Open Source movement, BSDI, Free BSD
  • hardware-independent Unix (unlike Solaris, Ultrix, SCO, etc)
  • teaching, self-teaching, hacker toy

Overall Architecture

• system calls: access to the OS from the application
• interrupts: access to the OS from the hardware
• separate address space for each process, one address space for kernel
• Unix-like file system, device files ("/dev"), configuration
• "/proc" file system for user/application access to kernel variables (kernel state)

Linux Networking

• three layers (two context switches) on receive:
  • interrupt handler
  • data link (ethernet, FDDI, slip, etc), IP, transport (TCP, UDP)
  • application
• two layers on send:
  • application, transport, IP, data link
  • interrupt handler

Comparison to Xinu

Comparison to X-kernel

Linux Organization

• Linux source code at http://www.ics.hawaii.edu/~esb/1999spring.ics651/linux (without SCSI or sound drivers, alpha, M68K, MIPS)
• include/linux/skbuff.h, net/core/skbuff.c: socket buffers
• include/net/sock.h: struct sock
• drivers/net/Space.c: device configuration
• drivers/net/slip.c: simple device
• drivers/net/eexpress.h: ethernet device

Linux Organization, continued

• net/core/dev.c: device management, reading, writing:
  • do_dev_queue_xmit: put a packet in a device queue, cause an interrupt
  • netif_rx: find the appropriate queue for an incoming packet (called by interrupt handler)
• net/ethernet/eth.c: ethernet header
• net/ipv4/*: IP, TCP, UDP
  • ip_input.c: read queue, dispatch packet
  • ip_output.c: fill in IP header fields
  • ip_masq*: IP masquerading (firewall)
  • tcp_input.c: TCP input processing
  • tcp.c: application-level TCP interface

UDP/IP/Ethernet packet receive

1. eexp_irq handles interrupt, calls eexp_hw_rx_pio to find out which buffers contain data, create skbuffers pointing to them, then calls netif_rx to queue the buffer and mark_bh to tell the kernel to execute net_bh
2. the kernel calls net_bh, which transmits any pending packets, removes packets from the queue, and delivers it to all the upcalls that handle this protocol type
3. the upcall for IP is ip_rcv, which after sanity checks upcalls the transport-level handler
4. the UDP upcall is udp_rcv, which finds the appropriate socket structure and calls udp_deliver which queues the packet
5. UDP receive is done by udp_recvmsg, which calls skb_recv_datagram to blocks if needed. udp_recvmsg then copies the datagram to user space and returns

UDP/IP/Ethernet packet send

1. udp_sendmsg calls udp_sendto calls udp_send after locking the socket
2. udp_send fills in the UDP and pseudo headers, and calls ip_build_xmit
3. ip_build_xmit fills in the IP header, copies the data to an skbuffer, and calls dev_queue_xmit, which enters a critical section in which it calls do_dev_queue_xmit
4. do_dev_queue_xmit puts a packet into a device queue and starts the hardware (dev->hard_start_xmit(skb, dev))