19: Networking. Networking Hardware. Mark Handley

Transcription

1 19: Networking Mark Handley Networking Hardware Lots of different hardware: Modem byte at a time, FDDI, SONET packet at a time ATM (including some DSL) 53-byte cell at a time Reality is that most networking is Internet-related, and most hardware presents a packet-based interface to the OS. is the canonical example networking interface. Everything else looks something like ethernet these days. 1

2 Original model is a shared broadcast-based media. CSMA-CD Try to sense carrier signal and avoid sending when someone else is. Detect collision and randomly backoff. MAC address: 48 bytes, (nominally) globally unique. Send to an address. Everyone on the LAN hears. Only the correct receiver actually receives the packet and passes it to the OS. host host host host host Modern ethernet is normally neither broadcast nor shared, but same API applies. Raw and Send Interrupts 1. Application prepares a packet to send. 2. Calls write (send) system call. 3. OS copies packet to kernel memory. 4. OS prepends ethernet header (14 bytes) to packet. May need to copy packet again, or use scatter/gather hardware. 5. OS programs DMA controller to send packet to Network Interface Card (NIC). Copy happens in background, controlled by DMA controller. 6. NIC sends packet on the wire. 7. NIC causes an interrupt when transmit is complete, or a failure occurs. 8. Sending process is unblocked (if it was blocked on transmit) 2

3 Raw and Receive Interrupts 1. Packet arrives at NIC. 2. Packet is destined for NIC s MAC address, or is broadcast, or is multicast (and NIC is listening) or NIC is in promiscuous mode. All other packets are silently ignored by NIC. 3. NIC transfers packet to host memory using pre-programmed DMA buffer. 4. NIC interrupts OS to say packet has arrived. 5. OS checks packet header fields to determine what to do with the packet. 6. OS copies the packet to user-space buffer Unblocks waiting process if it was blocked on receive. The Internet Protocol Stack Protocol stack is layered. Each layer provides a welldefined API to the layers above, and an expected use model to the layers below. is the common denominator understood by end-hosts and routers alike. WWW phone... SMTP HTTP RTP... UDP ethernet PPP CSMA async sonet... copper fiber radio... 3

4 Protocol Layering Link layers (eg ) are local to a particular link Routers look at headers to decide how to route a packet. provides reliability via retransmission, flow control, etc. Application using OS s API to do its job. HTTP HTTP ATM ATM Modem Modem Web Server Internet Router Internet Router Web Client / Stack: The OS s Role. Link layers (eg ) are local to a particular link Routers look at headers to decide how to route a packet. provides reliability via retransmission, flow control, etc. Application using OS s API to do its job. Application OS Kernel NIC hardware HTTP Management of buffered data, connection maintenance, retransmission of lost or corrupted packets, de-multiplex of received data. Host routing, filtering of received data, fragmentation and reassembly. ARP (->MAC address mapping), DMA buffer management 4

5 A Simple Example Computer A Computer B router In the sending host 1. Application-Programming Interface (API) Application in web client A requests connection with web server B. 2. Transmission Control Protocol () Creates Connection setup packet requests packet to be sent to B Packet Type = Connection Setup Empty 5

6 In the sending host 3. Internet Protocol () Creates packet with correct addresses. requests packet to be sent to router. Packet Encapsulation Packet Destination Address: B Source Address: A Protocol = In the sending host 4. Link ( MAC or ) Protocol Creates MAC frame Wait for Access to the line. MAC requests PHY to send each bit of the frame. Packet Encapsulation Destination Address: MAC R1 Source Address: MAC A Protocol = Packet 6

7 A Simple Example Computer A Computer B router Packet Packet travels over the first link In Router 5. Link ( MAC or ) Protocol Accept MAC frame, check address and CRC. Pass data to Protocol. Packet Destination Address: MAC R1 Source Address: MAC A Protocol = Decapsulation Packet 7

8 In Router 6. Internet Protocol () Use destination address to decide where to send packet next ( next-hop routing ). Request Link Protocol to transmit packet. Packet Destination Address: B Source Address: A Protocol = In Router 7. Link ( MAC or ) Protocol Creates MAC frame with new CRC. Wait for Access to the line. MAC requests PHY to send each bit of the frame. Packet Encapsulation Destination Address: MAC B Source Address: MAC R1 Protocol = Packet 8

9 A Simple Example Computer A Computer B router Packet Packet travels over the second link In the receiving host 8. Link ( MAC or ) Protocol Accept MAC frame, check address and CRC Pass data to Protocol. Packet Destination Address: MAC B Source Address: MAC R1 Protocol = Decapsulation Packet 9

10 In the receiving host 9. Internet Protocol () Verify address. Extract/decapsulate packet from packet. Pass packet to Protocol. Packet Decapsulation Destination Address: B Source Address: A Protocol = Packet In the receiving host 10. Transmission Control Protocol () Accepts Connection setup packet Establishes connection by sending Ack. 11. Application-Programming Interface (API) Application receives request for connection with A. Packet Type = Connection Setup Empty 10

11 Ugly Details Provides end-to-end addressing. Applications care about addresses. Hosts are configured with addresses. links hosts together on a LAN. Just plug it in. (MAC) addresses are part of the hardware. How does the OS know which ethernet address to put in a packet destined for a particular address? ARP: The Ugly Details How does the OS know which ethernet address to put in a packet destined for a particular address? ARP: Address Resolution Protocol. OS broadcasts a request onto the ethernet saying who has address ? Host replies saying That s me! I m at MAC address 08:00:20:77:e8:4f Pretty much any link layer will need some bootstrap mechanism similar to ARP to be provided by the OS because there s an abstraction boundary provided by the interface. 11

12 State OS must store state for every connection that is in progress. addresses and ports. Connection status. Current sequence numbers. Unacknowledged data at sender. Retransmit timer. Congestion state (window, RTT, RTT variance) State is stored in a Control Block (TCB). Efficient management of this state makes a big difference to OS performance. TCB Management Lookup speed vs complexity. On a busy server, can t afford to do an expensive lookup on every packet to find the TCB. Space management. Kernel memory is limited (TCB usually needs to be wired in memory). Tradeoff between lots of memory per TCB and lots of connections. Don t allocate TCB before connection handshake completes (DoS attack issue). 12

13 Making it Fast Can implement each layer as separate code, with rigid APIs between them. Performance will be poor. Buffering Copying the data. Cache locality. Context switching. High performance requires doing as much processing as possible in one pass in the interrupt handler without copying the data and with good cache locality. Interrupt mitigation techniques on both sender and receiver. Sockets API Basic networking API is Berkeley Sockets. Clearly derived from Unix everything is a file mentality, but also used on Windows. Call socket() bind() connect() listen() accept() write() read() close() Creates a communications endpoint and returns a file descriptor. Equivalent of open() for a file. Allocates a local address to a socket Establish a connection to a remote host Causes a socket to listen for incoming connections. Accepts an incoming connection. Sends data Receives data. Purpose Flushes buffers and closes the socket. 13