05-01-2010, 03:44 PM
[attachment=1058]
ABSTRACT
lwIP is an implementation of the TCP/IP protocol stack.. Interest for connecting small devices to existing network infrastructure such as global internet is steadily increasing. Such devices often has very limited CPU and memory resources and may not able to run an instance of TCP/IP protocol suite.
The focus of the lwIP stack is to reduce memory usage and code size, making lwIP suitable for use in small clients with very limited resources such as embedded systems. In order to reduce processing and memory demands, lwIP uses a tailor made API that does not require any data copying.
This report describes the design and main features that make lwIP to operate in minimal resource systems.
CHAPTER-1
1.1 INTRODUCTION
Over the last few years, the interest for connecting computers and computer supported devices to wireless networks has steadily increased. Computers are becoming more and more seamlessly integrated with everyday equipment and prices are dropping. At the same time wireless networking technologies, such as Bluetooth and IEEE 802.11b WLAN , are emerging. This gives rise to many new fascinating scenarios in areas such as health care, safety and security, transportation, and processing industry. Small devices such as sensors can be connected to an existing network infrastructure such as the global Internet, and monitored from anywhere.
The Internet technology has proven itself flexible enough to incorporate the changing network environments of the past few decades. While originally developed for low speed networks such as the ARPANET, the Internet technology today runs over a large spectrum of link technologies with vastly different characteristics in terms of bandwidth and bit error rate. It is highly advantageous to use the existing Internet technology in the wireless networks of tomorrow since a large amount of applications using the Internet technology have been developed. Also, the large connectivity of the global Internet is a strong incentive.
Since small devices such as sensors are often required to be physically small and inexpensive, an implementation of the Internet protocols will have to deal with having limited computing resources and memory. This report describes the design and implementation of a small TCP/IP stack called lwIP that is small enough to be used in minimal systems.
CHAPTER-2
OVERVIEW
As in many other TCP/IP implementations, the layered protocol design has served as a guide for the design of the implementation of lwIP. Each protocol is implemented as its own module, with a few functions acting as entry points into each protocol. Even though the protocols are implemented separately, some layer violations are made, as discussed above, in order to improve performance both in terms of processing speed and memory usage. For example, when verifying the checksum of an incoming TCP segment and when demultiplexing a segment, the source and destination IP addresses of the segment has to be known by the TCP module. Instead of passing these addresses to TCP by the means of a function call, the TCP module is aware of the structure of the IP header, and can therefore extract this information by itself.
lwIP consists of several modules. Apart from the modules implementing the TCP/IP protocols (IP, ICMP, UDP, and TCP) a number of support modules are implemented.
The support modules consists of :-
¢ The operating system emulation layer (described in Chapter3)
¢ The buffer and memory management subsystems
(described in Chapter 4)
¢ Network interface functions (described in Chapter 5)
¢ Functions for computing Internet checksum (Chapter 6)
¢ An abstract API (described in Chapter 8 )
2.1 PROTOCOL LAYERING
The protocols in the TCP/IP suite are designed in a layered fashion, where each protocol layer solves a separate part of the communication problem. This layering can serve as a guide for designing the implementation of the protocols, in that each protocol can be implemented separately from the other. Implementing the protocols in a strictly layered way can however, lead to a situation where the communication overhead between the protocol layers degrades the overall performance. To overcome these problems, certain internal aspects of a protocol can be made known to other protocols. Care must be taken so that only the important information is shared among the layers.
Most TCP/IP implementations keep a strict division between the application layer and the lower protocol layers, whereas the lower layers can be more or less interleaved. In most operating systems, the lower layer protocols are implemented as a part of the operating system kernel with entry points for communication with the application layer process. The application program is presented with an abstract view of the TCP/IP implementation, where network communication differs only very little from inter-process communication or file I/O. The implications of this is that since the application program is unaware of the buffer mechanisms used by the lower layers, it cannot utilize this information to, e.g., reuse buffers with frequently used data. Also, when the application sends data, this data has to be copied from the application process' memory space into internal buffers before being processed by the network code.
The operating systems used in minimal systems such as the target system of lwIP most often do not maintain a strict protection barrier between the kernel and the application processes. This allows using a more relaxed scheme for communication between the application and the lower layer protocols by the means of shared memory. In particular, the application layer can be made aware of the buffer handling mechanisms used by the lower layers. Therefore, the application can more efficiently reuse buffers. Also, since the application process can use the same memory as the networking code the application can read and write directly to the internal buffers, thus saving the expense of performing a copy.
2.2 PROCESS MODEL
The process model of a protocol implementation describes in which way the system has been divided into different processes. One process model that has been used to implement communication protocols is to let each protocol run as a stand alone process. With this model, a strict protocol layering is enforced, and the communication points between the protocols must be strictly defined.
While this approach has its advantages such as protocols can be added at runtime, understanding the code and debugging is generally easier, there are also disadvantages. The strict layering is not, as described earlier, always the best way to implement protocols. Also, and more important, for each layer crossed, a context switch must be made. For an incoming TCP segment this would mean three context switches, from the device driver for the network interface, to the IP process, to the TCP process and finally to the application process. In most operating systems a context switch is fairly expensive.
Another common approach is to let the communication protocols reside in the kernel of the operating system. In the case of a kernel implementation of the communication protocols, the application processes communicate with the protocols through system calls. The communication protocols are not strictly divided from each other but may use the techniques of crossing the protocol layering.
lwIP uses a process model in which all protocols reside in a single process and are thus separated from the operating system kernel. Application programs may either reside in the lwIP process, or be in separate processes. Communication between the TCP/IP stack and the application programs are done either by function calls for the case where the application program shares a process with lwIP, or by the means of a more abstract API.
Having lwIP implemented as a user space process rather than in the operating system kernel has both its advantages and disadvantages. The main advantage of having lwIP as a process is that is portable across different operating systems. Since lwIP is designed to run in small operating systems that generally do not support neither swapping out processes not virtual memory, the delay caused by having to wait for disk activity if part of the lwIP process is swapped or paged out to disk will not be a problem. The problem of having to wait for a scheduling quantum before getting a chance to service requests still is a problem however, but there is nothing in the design of lwIP that precludes it from later being implemented in an operating system kernel.
CHAPTER-3
THE OPERATING SYSTEM EMULATION LAYER
In order to make lwIP portable, operating system specific function calls and data structures are not used directly in the code. Instead, when such functions are needed the operating system emulation layer is used. The operating system emulation layer provides a uniform interface to operating system services such as timers, process synchronization, and message passing mechanisms. In principle, when porting lwIP to other operating systems only an implementation of the operating system emulation layer for that particular operating system is needed.
The operating system emulation layer provides a timer functionality that is used by TCP. The timers provided by the operating system emulation layer are one-shot timers with a granularity of at least 200 ms that calls a registered function when the time-out occurs.
The only process synchronization mechanism provided is semaphores. Even if semaphores are not available in the underlying operating system they can be emulated by other synchronization primitives such as conditional variables or locks.
The message passing is done through a simple mechanism which uses an abstraction called mailboxes. A mailbox has two operations: post and fetch. The post operation will not block the process; rather, messages posted to a mailbox are queued by the operating system emulation layer until another process fetches them. Even if the underlying operating system does not have native support for the mailbox mechanism, they are easily implemented using semaphores.
CHAPTER-4
BUFFER AND MEMORY MANAGEMENT
The memory and buffer management system in a communication system must be prepared to accommodate buffers of very varying sizes, ranging from buffers containing full-sized TCP segments with several hundred bytes worth of data to short ICMP echo replies consisting of only a few bytes. Also, in order to avoid copying it should be possible to let the data content of the buffers reside in memory that is not managed by the networking subsystem, such as application memory or ROM.
4.1 Packet buffers | pbufs
A pbuf is lwIP's internal representation of a packet, and is designed for the special needs of the minimal stack. Pbufs are similar to the mbufs used in the BSD implementations. The pbuf structure has support both for allocating dynamic memory to hold packet contents, and for letting packet data reside in static memory. Pbufs can be linked together in a list, called a pbuf chain so that a packet may span over several pbufs.
Pbufs are of three types, PBUF RAM, PBUF ROM, and PBUF POOL. The pbuf shown in Figure 1 represents the PBUF RAM type, and has the packet data stored in memory managed by the pbuf subsystem. The pbuf in Figure 2 is an example of a chained pbuf, where the first pbuf in the chain is of the PBUF RAM type, and the second is of the PBUF ROM type, which means that it has the data located in memory not managed by the pbuf system. The third type of pbuf, PBUF POOL, is shown in Figure 3 and consists of fixed size pbufs allocated from a pool of fixed size pbufs. A pbuf chain may consist of multiple types of pbufs.
The three types have different uses. Pbufs of type PBUF POOL are mainly used by network device drivers since the operation of allocating a single pbuf is fast and is therefore suitable for use in an interrupt handler. PBUF ROM pbufs are used when an application sends data that is located in memory managed by the application. This data may not be modified after the pbuf has been handed over to the TCP/IP stack and therefore this pbuf type main use is when the data is located in ROM (hence the name PBUF ROM). Headers that are prepended to the data in a PBUF ROM pbuf are stored in a PBUF RAM pbuf that is chained to the front of of the PBUF ROM pbuf, as in Figure 2
Pbufs of the PBUF RAM type are also used when an application sends data that is dynamically generated. In this case, the pbuf system allocates memory not only for the application data, but also for the headers that will be prepended to the data. This is seen in Figure 1. The pbuf system cannot know in advance what headers will be prepended to the data and assumes the worst case.The size of the headers is configurable at compile time.
In essence, incoming pbufs are of type PBUF POOL and outgoing pbufs are of the PBUF ROM or PBUF RAM types.
The internal structure of a pbuf can be seen in the Figures 1 through 3. The pbuf structure consists of two pointers, two length fields, flags field, and a reference count. The next field is a pointer to the next pbuf in case of a pbuf chain. The payload pointer points to the start of the data in the pbuf. The len field contains the length of the data contents of the pbuf. The tot len field contains the sum of the length of the current pbuf and all len fields of following pbufs in the pbuf chain. In other words, the tot len field is the sum of the len field and the value of the tot len field in the following pbuf in the pbuf chain. The flags field indicates the type of the pbuf and the ref field -contains a reference count. The next and payload fields are native pointers and the size of those varies depending on the processor architecture used. The two length fields are 16 bit unsigned integers and the flags and ref fields are 4 bit wide. The total size of the pbuf structure depends on the size of a pointer in the processor architecture being used and on the smallest alignment possible for the processor architecture. On an architecture with 32 bit pointers and 4 byte alignment, the total size is 16 bytes and on an architecture with 16 bit pointers and 1 byte alignment, the size is 9 bytes.
4.2 Memory management
The memory manager supporting the pbuf scheme is very simple. It handles allocations and deallocations of contiguous regions of memory and can shrink the size of a previously allocated memory block. The memory manager uses a dedicated portion of the total memory in the system. This ensures that the networking system does not use all of the available memory, and that the operation of other programs is not disturbed if the networking system has used all of it's memory.
Internally, the memory manager keeps track of the allocated memory by placing a small structure on top of each allocated memory block. This structure holds two pointers to the next and previous allocation block in memory. It also has a used flag which indicates whether the allocation block is allocated or not.
Memory is allocated by searching the memory for an unused allocation block that is large enough for the requested allocation. The first-fit principle is used so that the first block that is large enough is used. When an allocation block is deallocated, the used flag is set to zero. In order to prevent fragmentation, the used flag of the next and previous allocation blocks are checked. If any of them are unused, the blocks are combined into one larger unused block.
CHAPTER-5
NETWORK INTERFACES
In lwIP device drivers for physical network hardware are represented by a network interface structure. The network interfaces are kept on a global linked list, which is linked by the next pointer in the structure. Each network interface has a name, stored in the name field . This two letter name identifies the kind of device driver used for the network interface and is only used when the interface is configured by a human operator at runtime. The name is set by the device driver and should reflect the kind of hardware that is represented by the network interface. For example, a network interface for a Bluetooth driver might have the name bt and a network interface for IEEE 802.11b WLAN hardware could have the name wl. Since the names not necessarily are unique, the num field is used to distinguish different network interfaces of the same kind.
Figure 4. The memory allocation structure.
The three IP addresses ip addr, netmask and gw are used by the IP layer when sending and receiving packets, and their use is described in the next section. It is not possible to configure a network interface with more than one IP address. Rather, one network interface would have to be created for each IP address.
The input pointer points to the function the device driver should call when a packet has been received.A network interface is connected to a device driver through the output pointer. This pointer points to a function in the device driver that transmits a packet on the physical network and it is called by the IP layer when a packet is to be sent. This field is filled by the initialization
function of the device driver. The third argument to the output function, ipaddr, is the IP address of the host that should receive the actual link layer frame. It does not have to be the same as the destination address of the IP packet. In particular, when sending an IP packet to a host that is not on the local network, the link level frame will be sent to a router on the network. In this case, the IP address given to the output function will be the IP address of the router.
Finally, the state pointer points to device driver specific state for the network interface and is set by the device driver.
CHAPTER-6
IMPLEMENTATION OF IP,UDP,TCP
6.1 IP Processing
lwIP implements only the most basic functionality of IP. It can send, receive and forward packets, but cannot send or receive fragmented IP packets nor handle packets with IP options. For most applications this does not pose any problems.
6.1.1 Receiving packets
For incoming IP packets, processing begins when the ip input() function is called by a network device driver. Here, the initial sanity checking of the IP version field and the header length is done, as well as computing and checking the header checksum. It is expected that the stack will not receive any IP fragments since the proxy is assumed to reassemble any fragmented packets, thus any packet that is an IP fragment is silently discarded. Packets carrying IP options are also assumed to be handled by the proxy, and are dropped.
Next, the function checks the destination address with the IP addresses of the network interfaces to determine if the packet was destined for the host. The network interfaces are ordered in a linked list, and it is searched linearly. The number of network interfaces is expected to be small so a more sophisticated search strategy than a linear search has not been implemented.
If the incoming packet is found to be destined for this host, the protocol field is used to decide to which higher level protocol the packet should be passed to.
6.1.2 Sending packets
An outgoing packet is handled by the function ip output(), which uses the function ip route() to find the appropriate network interface to transmit the packet on. When the outgoing network interface is determined, the packet is passed to ip output if() which takes the outgoing network interface as an argument. Here, all IP header fields are filled in and the IP header checksum is computed. The source and destination addresses of the IP packet is passed as an argument to ip output if(). The source address may be left out, however, and in this case the IP address of the outgoing network interface is used as the source IP address of the packet.
The ip route() function finds the appropriate network interface by linearly searching the list of network interfaces. During the search the destination IP address of the IP packet is masked with the netmask of the network interface. If the masked destination address is equal to the masked IP address of the interface, the interface is chosen. If no match is found, a default network interface is used. The default network interface is configured either at boot-up or at runtime by a human operator. If the network address of the default interface does not match the destination IP address, the gw field in the network interface structure is chosen as the destination IP address of the link level frame. (Notice that the destination address of the IP packet and the destination address of the link level frame will be different in this case.) This primitive form of routing glosses over the fact that a network might have many routers attached to it. For the most basic case, where a local network only has one router, this works however.
Since the transport layer protocols UDP and TCP need to have the destination IP address when computing the transport layer checksum, the outgoing network interface must in some cases be determined before the packet is passed to the IP layer. This is done by letting the transport layer functions call the ip route() function directly, and since the outgoing network interface is known already when the packet reaches the IP layer, there is no need to search the network interface list again. Instead, those protocols call the ip output if() function directly. Since this function takes a network interface as an argument, the search for an outgoing interface is avoided.
Human configuration of lwIP during runtime requires an application program that is able to configure the stack. Such a program is not included in lwIP.
6.1.3 Forwarding packets
If none of the network interfaces has the same IP address as an incoming packet's destination address, the packet should be forwarded. This is done by the function ip forward(). Here, the TTL field is decreased and if it reaches zero, an ICMP error message is sent to the original sender of the IP packet and the packet is discarded. Since the IP header is changed, the IP header checksum needs to be adjusted. The is no need to recompute the entire checksum, however, since simple arithmetic can be used to adjust the original IP checksum.
Finally, the packet is forwarded to the appropriate network interface. The algorithm used to find the appropriate network interface is the same that is used when sending IP packets.
6.1.4 ICMP processing
ICMP processing is fairly simple. ICMP packets received by ip input() are handed over to icmp input(), which decodes the ICMP header and takes the appropriate action. Some ICMP messages are passed to upper layer protocols and those are taken care of by special functions in the transport layer. ICMP destination unreachable messages can be sent by transport layer protocols, in particular by UDP, and the function icmp dest unreach() is used for this.
Using ICMP ECHO messages to probe a network is widely used, and therefore ICMP echo processing is optimized for performance. The actual processing takes place in icmp input(), and consists of swapping the IP destination and source addresses of the incoming packet, change the ICMP type to echo reply and adjust the ICMP checksum. The packet is then passed back to the IP layer for transmission.
6.2 UDP processing
UDP is a simple protocol used for demultiplexing packets between different processes. The state for each UDP session is kept in a PCB structure . The UDP PCBs are kept on a linked list which is searched for a match when a UDP datagram arrives.
The UDP PCB structure contains a pointer to the next PCB in the global linked list of UDP PCBs. A UDP session is defined by the IP addresses and port numbers of the end-points and these are stored in the local ip, dest ip, local port and dest port fields. The flags field indicates what UDP checksum policy that should be used for this session. This can be either to switch UDP checksumming off completely, or to use UDP Lite in which the checksum covers only parts of the datagram. If UDP Lite is used, the chksum len field specifies how much of the datagram that should be checksummed.
The last two arguments recv and recv arg are used when a datagram is received in the session specified by the PCB. The function pointed to by recv is called when a datagram is received.
Due to the simplicity of UDP, the input and output processing is equally simple and follows a fairly straight line (Figure 7 ). To send data, the application program calls udp send() which calls upon udp output(). Here the the necessary checksumming is done and UDP header fields are filled. Since the checksum includes the IP source address of the IP packet, the function ip route() is in some cases called to find the network interface to which the packet is to be transmitted. The IP address of this network interface is used as the source IP address of the packet. Finally, the packet is turned over to ip output if() for transmission.
fig 7 UDP processing
When a UDP datagram arrives, the IP layer calls the udp input() function. Here, if check-summing should be used in the session, the UDP checksum is checked and the datagram is demultiplexed. When the corresponding UDP PCB is found, the recv function is called.
6.3 TCP processing
TCP is a transport layer protocol that provides a reliable byte stream service to the application layer. TCP is more complex than the other protocols described here, and the TCP code constitutes50% of the total code size of lwIP.
6.3.1 Overview
The basic TCP processing (Figure 8 ) is divided into six functions; the functions tcp input(),tcp process(), and tcp receive() which are related to TCP input processing, and tcp write(),tcp enqueue(), and tcp output() which deals with output processing.
Fig 8 TCP processing
When an application wants to send TCP data, tcp write() is called. The function tcp write() passes control to tcp enqueue() which will break the data into appropriate sized TCP segmentsif necessary and put the segments on the transmission queue for the connection. The function tcp output() will then check if it is possible to send the data, i.e., if there is enough space in the receiver's window and if the congestion window is large enough and if so, sends the data using ip route() and ip output if().
Input processing begins when ip input() after verifying the IP header hands over a TCP segment to tcp input(). In this function the initial sanity checks (i.e., checksumming and TCP options parsing) are done as well as deciding to which TCP connection the segment belongs. The segment is then processed by tcp process(), which implements the TCP state machine, and any necessary state transitions are made. The function tcp receive() will be called if the connection is in a state to accept data from the network. If so, tcp receive() will pass the segment up to an application program. If the segment constitutes an ACK for unacknowledged (thus previously buffered) data, the data is removed from the buffers and its memory is reclaimed. Also, if an ACK for data was received the receiver might be willing to accept more data and therefore tcp output() is called.
6.3.2 Data structures
The data structures used in the implementation of TCP are kept small due to the memory constraints in the minimal system for which lwIP is intended. There is a tradeoff between the complexity of the data structures and the complexity of the code that uses the data structures, and in this case the code complexity has been sacrificed in order to keep the size of the data structures small.
The TCP PCB is fairly large. Since TCP connections in the LISTEN and TIME-WAIT needs to keep less state information than connections in other states, a smaller PCB data structure is used for those connections. This data structure is overlaid with the full PCB structure, and the ordering of the items in the PCB structure is therefore somewhat awkward.
The TCP PCBs are kept on a linked list, and the next pointer links the PCB list together. The state variable contains the current TCP state of the connection. Next, the IP addresses and port numbers which identify the connection are stored. The mss variable contains the maximum segment size allowed for the connection.
The rcv nxt and rcv wnd fields are used when receiving data. The rcv nxt field contains the next sequence number expected from the remote end and is thus used when sending ACKs to the remote host. The receiver's window is kept in rcv wnd and this is advertised in outgoing TCP segments. The field tmr is used as a timer for connections that should be removed after a certain amount of time, such as connections in the TIME-WAIT state. The maximum segment size allowed on the connection is stored in the mss field. The flags field contains additional state information of the connection, such as whether the connection is in fast recovery or if a delayed ACK should be send.
The fields rttest, rtseq, sa, and sv are used for the round-trip time estimation. The sequence number of the segment that is used for estimating the round-trip time is stored in rtseq and the time this segment was sent is stored in rttest. The average round-trip time and the round-trip time variance is stored in sa and sv. These variables are used when calculating the retransmission time-out which is stored in the rto field.
The two fields lastack and dupacks are used in the implementation of fast retransmit and fast recovery. The lastack field contains the sequence number acknowledged by the last ACK received and dupacks contains a count of how many ACKs that has been received for the sequence number in lastack. The current congestion window for the connection is stored in the cwnd field and the slow start threshold is kept in ssthresh.
The six fields snd ack, snd nxt, snd wnd, snd wl1, snd wl2 and snd lbb are used when sending data. The highest sequence number acknowledged by the receiver is stored in snd ack and the next sequence number to send is kept in snd nxt. The receiver's advertised window is held in snd wnd and the two fields snd wl1 and snd wl2 are used when updating snd wnd. The snd lbb field contains the sequence number of the last byte queued for transmission.
The function pointer recv and recv arg are used when passing received data to the application layer. The three queues unsent, unacked and ooseq are used when sending and receiving data. Data that has been received from the application but has not been sent is queued in unsent and data that has been sent but not yet acknowledged by the remote host is held in unacked. Received data that is out of sequence is buffered in ooseq.
struct tcp_seg {
struct tcp_seg *next;
u16_t len;
struct pbuf *p;
struct tcp_hdr *tcphdr;
void *data;
u16_t rtime;
};
Figure 9. The tcp seg structure
The tcp seg structure in Figure 9 is the internal representation of a TCP segment. This structure starts with a next pointer which is used for linking when queuing segments. The len field contains the length of the segment in TCP terms. This means that the len field for a data segment will contain the length of the data in the segment, and the len field for an empty segment with the SYN or FIN flags set will be 1. The pbuf p is the buffer containing the actual segment and the tcphdr and data pointers points to the TCP header and the data in the segment, respectively.
For outgoing segments, the rtime field is used for the retransmission time-out of this segment. Since incoming segments will not need to be retransmitted, this field is not needed and memory for this field is not allocated for incoming segments.
6.3.3 Sequence number calculations
The TCP sequence numbers that are used to enumerate the bytes in the TCP byte stream are unsigned 32 bit quantities, hence in the range [0; 232 ¡ 1]. Since the number of bytes sent in a TCP connection might be more than the number of 32-bit combinations, the sequence numbers are calculated modulo 232. This means that ordinary comparison operators cannot be used withTCP sequence numbers. The modified comparison operators, called <seq and >seq, are defined by the relations
s <seq t , s ¡ t < 0
and
s >seq t , s ¡ t > 0;
where s and t are TCP sequence numbers. The comparison operators for ¢ and ¸ are defined equivalently. The comparison operators are defined as C macros in the header file.
6.3.4 Queuing and transmitting data
Data that is to be sent is divided into appropriate sized chunks and given sequence numbers by the tcp enqueue() function. Here, the data is packeted into pbufs and enclosed in a tcp seg struct The TCP header is build in the pbuf, and filled in with all fields except the acknowledgment number, ackno, and the advertised window, wnd. These fields can change during the queuing time of the segment and are therefore set by tcp output() which does the actual transmission of the segment. After the segments are built, they are queued on the unsent list in the PCB. The tcp enqueue() function tries to fill each segment with a maximum segment size worth of data and when an under-full segment is found at the end of the unsent queue, this segment is appended with the new data using the pbuf chaining functionality.
After tcp enqueue() has formatted and queued the segments, tcp output() is called. It checks if there is any room in the current window for any more data. The current window is computed by taking the maximum of the congestion window and the advertised receiver's window. Next, it fills in the fields of the TCP header that was not filled in by tcp enqueue() and transmits the segment using ip route() and ip output if(). After transmission the segment is put on the unacked list, on which it stays until an ACK for the segment has been received.
When a segment is on the unacked list, it is also timed for retransmission . When a segment is retransmitted the TCP and IP headers of the original segment is kept and only very little changes has to be made to the TCP header. The ackno and wnd fields of the TCP header are set to the current values since we could have received data during the time between the original transmission of the segment and the retransmission. This changes only two 16-bit words in the header and the whole TCP checksum does not have to be recomputed since simple arithmetic can be used to update the checksum. The IP layer has already added the IP header when the segment was originally transmitted and there is no reason to change it. Thus a retransmission does not require any recomputation of the IP header checksum.
6.3.4.1 Silly window avoidance
The Silly Window Syndrome (SWS) is a TCP phenomena that can lead to very bad performance. SWS occurs when a TCP receiver advertises a small window and the TCP sender immediately sends data to fill the window. When this small segment is acknowledged the window is opened again by a small amount and sender will again send a small segment to fill the window. This leads to a situation where the TCP stream consists of very small segments. In order to avoid SWS both the sender and the receiver must try to avoid this situation. The receiver must not advertise small window updates and the sender must not send small segments when only a small window is offered.
In lwIP SWS is naturally avoided at the sender since TCP segments are constructed and queued without knowledge of the advertised receiver's window. In a large transfer the output queue will consist of maximum sized segments. This means that if a TCP receiver advertises a small window, the sender will not send the first segment on the queue since it is larger than the advertised window. Instead, it will wait until the window is large enough for a maximum sized segment.
When acting as a TCP receiver, lwIP will not advertise a receiver's window that is smaller than the maximum segment size of the connection.
6.3.5 Receiving segments
6.3.5.1 Demultiplexing
When TCP segments arrive at the tcp input() function, they are demultiplexed between the TCP PCBs. The demultiplexing key is the source and destination IP addresses and the TCP port numbers. There are two types of PCBs that must be distinguished when demultiplexing a segment; those that correspond to open connections and those that correspond to connections that are half open. Half open connections are those that are in the LISTEN state and only have the local TCP port number specified and optionally the local IP address, whereas open connections have the both IP addresses and both port numbers specified.
Many TCP implementations, use a technique where a linked list of PCBs with a single entry cache is used. The rationale behind this is that most TCP connections constitute bulk transfers which typically show a large amount locality, resulting in a high cache hit ratio. Other caching schemes include keeping two one entry caches, one for the PCB corresponding to the last packet that was sent and one for the PCB of the last packet received . An alternative scheme to exploit locality can be done by moving the most recently used PCB to the front of the list. Both methods have been shown to outperform the one entry cache scheme.
In lwIP, whenever a PCB match is found when demultiplexing a segment, the PCB is moved to the front of the list of PCBs. PCBs for connections in the LISTEN state are not moved to the front however, since such connections are not expected to receive segments as often as connections that are in a state in which they receive data.
6.3.5.2 Receiving data
The actual processing of incoming segments is made in the function tcp receive(). The acknowledgment number of the segment is compared with the segments on the unacked queue of the connection. If the acknowledgment number is higher than the sequence number of a segment on the unacked queue, that segment is removed from the queue and the allocated memory for the segment is deallocated.
An incoming segment is out of sequence if the sequence number of the segment is higher than the rcv nxt variable in the PCB. Out of sequence segments are queued on the ooseq queue in the PCB. If the sequence number of the incoming segment is equal to rcv nxt, the segment is delivered to the upper layer by calling the recv function in the PCB and rcv nxt is increased by the length of the incoming segment. Since the reception of an in-sequence segment might mean that a previously received out of sequence segment now is the next segment expected, the ooseq queued is checked. If it contains a segment with sequence number equal to rcv nxt, this segment is delivered to the application by a call to to recv function and rcv nxt is updated. This process continues until either the ooseq queue is empty or the next segment on ooseq is out of sequence.
6.3.6 Accepting new connections
TCP connections that are in the LISTEN state, i.e., that are passively open, are ready to accept new connections from a remote host. For those connections a new TCP PCB is created and must be passed to the application program that opened the initial listening TCP connection. In lwIP this is done by letting the application register a callback function that is to be called when a new connection has been established.
When a connection in the LISTEN state receives a TCP segment with the SYN flag set, a new connection is created and a segment with the SYN and ACK flags are sent in response to the SYN segment. The connection then enters the SYN-RCVD state and waits for an acknowledgment for the sent SYN segment. When the acknowledgment arrives, the connection enters the ESTABLISHED state, and the accept function (the accept field in the PCB structure is called.
6.3.7 Fast retransmit
Fast retransmit and fast recovery is implemented in lwIP by keeping track of the last sequence number acknowledged. If another acknowledgment for the same sequence number is received, the dupacks counter in the TCP PCB is increased. When dupacks reaches three, the first segment on the unacked queue is retransmitted and fast recovery is initialized. The implementation of fast recovery follows the steps laid out in . Whenever an ACK for new data is received, the dupacks counter is reset to zero.
6.3.8 Timers
As in the the BSD TCP implementation, lwIP uses two periodical timers that goes off every 200 ms and 500 ms. Those two timers are then used to implement more complex logical timers such as the retransmission timers, the TIME-WAIT timer and the delayed ACK timer.
The fine grained timer, tcp timer fine() goes through every TCP PCB checking if there are any delayed ACKs that should be sent, as indicated by the flag field in the tcp pcb structure . If the delayed ACK flag is set, an empty TCP acknowledgment segment is sent and the flag is cleared.
The coarse grained timer, implemented in tcp timer coarse(), also scans the PCB list. For every PCB, the list of unacknowledged segments (, is traversed, and the rtime variable is increased. If rtime becomes larger than the current retransmission time-out as given by the rto variable in the PCB, the segment is retransmitted and the retransmission time-out is doubled. A segment is retransmitted only if allowed by the values of the congestion window and the advertised receiver's window. After retransmission, the congestion window is set to one maximum segment size, the slow start threshold is set to half of the effective window size, and slow start is initiated on the connection.
For connections that are in TIME-WAIT, the coarse grained timer also increases the tmr field in the PCB structure. When this timer reaches the 2*MSL threshold, the connection is removed.The coarse grained timer also increases a global TCP clock, tcp ticks. This clock is used for round-trip time estimation and retransmission time-outs.
6.3.9 Round-trip time estimation
The round-trip time estimation is a critical part of TCP since the estimated round-trip time is used when determining a suitable retransmission time-out. In lwIP round-trip times measurements are taken in a fashion similar to the BSD implementations. Round-trip times are measured once per round-trip and the smoothing function described in is used for the calculation of a suitable retransmission time-out.
The TCP PCB variable rtseq hold the sequence number of the segment for which the roundtrip time is measured. The rttest variable in the PCB holds the value of tcp ticks when the segment was first transmitted. When an ACK for a sequence number equal to or larger than rtseq is received, the round-trip time is measured by subtracting rttest from tcp ticks. If a retransmission occurred during the round-trip time measurement, no measurement is taken.
6.3.10 Congestion control
The implementation of congestion control is surprisingly simple and consists of a few lines of code in the output and input code. When an ACK for new data is received the congestion window, cwnd, is increased either by one maximum segment size or by mss2/cwnd, depending on whether the connection is in slow start or congestion avoidance. When sending data the minimum value of the receiver's advertised window and the congestion window is used to determine how much data that can be sent in each window.
CHAPTER-7
INTERFACING THE STACK
There are two ways for using the services provided by the TCP/IP stack; either by calling the functions in the TCP and UDP modules directly, or to use the lwIP API presented in the next section.
The TCP and UDP modules provide a rudimentary interface to the networking services. The interface is based on callbacks and an application program that uses the interface can therefore not operate in a sequential manner. This makes the application program harder to program and the application code is harder to understand. In order to receive data the application program registers a callback function with the stack. The callback function is associated with a particular connection and when a packet arrives in the connection, the callback function is called by the stack.
Furthermore, an application program that interfaces the TCP and UDP modules directly has to (at least partially) reside in the same process as the TCP/IP stack. This is due to the fact that a callback function cannot be called across a process boundary. This has both advantages and disadvantages. One advantage is that since the application program and the TCP/IP stack are in the same process, no context switching will be made when sending or receiving packets. The main disadvantage is that the application program cannot involve itself in any long running computations since TCP/IP processing cannot occur in parallel with the computation, thus degrading communication performance. This can be overcome by splitting the application into two parts, one part dealing with the communication and one part dealing with the computation. The part doing the communication would then reside in the TCP/IP process and the computationally heavy part would be a separate process. The lwIP API presented in the next section provides a structured way to divide the application in such a way.
CHAPTER-8
APPLICATION PROGRAM INTERFACE
Due to the high level of abstraction provided by the BSD socket API, it is unsuitable for use in a minimal TCP/IPimplementation. In particular, BSD sockets require data that is to be sent to be copied from the application program to internal buffers in the TCP/IP stack. The reason for copying the data is that the application and the TCP/IP stack usually reside in different protection domains. In most cases the application program is a user process and the TCP/IP stack resides in the operating system kernel. By avoiding the extra copy, the performance of the API can be greatly improved . Also, in order to make a copy, extra memory needs to be allocated for the copy, effectively doubling the amount of memory used per packet.
The lwIP API was designed for lwIP and utilizes knowledge of the internal structure of lwIP to achieve effectiveness. The lwIP API is very similar to the BSD API, but operates at a slightly lower level. The API does not require that data is copied between the application program and the TCP/IP stack, since the application program can manipulate the internal buffers directly.
.
8.1 BASIC CONCEPTS
From the application's point of view, data handling in the BSD socket API is done in continuous memory regions. This is convenient for the application programmer since manipulation of data in application programs is usually done in such continuous memory chunks. Using this type of mechanism with lwIP would not be advantageous, since lwIP usually handles data in buffers where the data is partitioned into smaller chunks of memory. Thus the data would have to be copied into a continuous memory area before being passed to the application. This would waste both processing time and memory since. Therefore, the lwIP API allows the application program to manipulate data directly in the partitioned buffers in order to avoid the extra copy.
The lwIP API uses a connection abstraction similar to that of the BSD socket API. There are very noticeable differences however; where an application program using the BSD socket API need not be aware of the distinction between an ordinary file and a network connection, an application program using the lwIP API has to be aware of the fact that it is using a network connection. Network data is received in the form of buffers where the data is partitioned into smaller chunks of memory. Since many applications wants to manipulate data in a continuous memory region, a convenience function for copying the data from a fragmented buffer to continuous memory exists.
Sending data is done differently depending on whether the data should be sent over a TCP connection or as UDP datagrams. For TCP, data is sent by passing the output function a pointer to a continuous memory region. The TCP/IP stack will partition the data into appropriately sized packets and queue them for transmission. When sending UDP datagrams, the application program will to explicitly allocate a buffer and fill it with data. The TCP/IP stack will send the datagram immediately when the output function is called.
8.2 IMPLEMENTATION OF THE API
The implementation of the API is divided into two parts, due to the process model of the TCP/IP stack. As shown in Figure 10 , parts of the API is implemented as a library linked to the application program, and parts are implemented in the TCP/IP process. The two parts communicate using the interprocess communication (IPC) mechanisms provided by the operating system emulation layer. The current implementation uses the following three IPC mechanisms:
¢ shared memory
¢ message passing
¢ semaphores.
While these IPC types are supported by the operating system layer, they need not be directly supported by the underlying operating system. For operating systems that do not natively support them, the operating system emulation layer emulates them.
Fig 10 Division of API implementation
The general design principle used is to let as much work as possible be done within the application process rather than in the TCP/IP process. This is important since all processes use the TCP/IP process for their TCP/IP communication. Keeping down the code footprint of the part of the API that is linked with the applications is not as important. This code can be shared among the processes, and even if shared libraries are not supported by the operating system, the code is stored in ROM. Embedded systems usually carry fairly large amounts of ROM, whereas processing power is scarce.
The buffer management is located in the library part of the API implementation. Buffers are created, copied and deallocated in the application process. Shared memory is used to pass the buffers between the application process and the TCP/IP process. The buffer data type used in communication with the application program is an abstraction of the pbuf data type.
Buffers carrying referenced memory, as opposed to allocated memory, is also passed using shared memory. For this to work, is has to be possible to share the referenced memory between the processes. The operating systems used in embedded systems for which lwIP is intended usually do not implement any form of memory protection, so this will not be a problem.
The functions that handle network connections are implemented in the part of the API implementation that resides in the TCP/IP process. The API functions in the part of the API that runs in the application process will pass a message using a simple communication protocol to the API implementation in the TCP/IP process. The message includes the type of operation that should be carried out and any arguments for the operation. The operation is carried out by the API implementation in the TCP/IP process and the return value is sent to the application process by message passing.
8.3 PERFORMANCE ANALYSIS
The performance of lwIP in terms of RAM usage or code effciency have not been formally tested and this has been noted in future work. Simple tests have been conducted, however, and these have shown that lwIP running a simple HTTP/1.0 server is able to serve at least 10 simultaneous requests for web pages while consuming less than 4 kilobytes of RAM. In those tests,only the memory used by the protocols, buffering system, and the application program has been taken into account.
SUMMARY
The small TCP/IP stack
lwIP, the small TCP/IP stack, has been shown to have a small code size. It is designed for systems with little RAM and simple tests have shown that it can operate in an environment with very little RAM available. Operating system dependencies are moved into a separate module to simplify porting of lwIP to other operating systems. The TCP/IP stack has support for TCP, UDP, ICMP, and IP. The TCP layer is fully functional and implements congestion control, round-trip time estimation, fast retransmit, and fast recovery. The IP layer has support for IP forwarding and can be used as a simple router.
REFERENCES
¢ Adam Dunkels, lightweight IP: An overview.Swedish Institute of Computer Science
¢ B. Ahlgren, M. BjÄorkman, and K. Moldeklev. The performance of a no-copy api for communication subsystems.
¢ M. Allman, V. Paxson, and W. Stevens. TCP congestion control. RFC 2581, Internet Engineering Task Force.
¢ D.D.Clark. Modularity and efficiency in protocol implementation. RFC 817, Internet Engineering Task Force.
¢ Paul E. McKenney and Ken F. Dove. Efficient demultiplexing of incoming TCP packets.
¢ T. Mallory and A. Kullberg. Incremental updating of the internet checksum. RFC 1141, Internet Engineering Task Force.
CONTENTS
Chapter 1
1.1 Introduction 01
Chapter 2
Overview 0
2.1 Protocol Layering 02
2.2 Process Model 03
Chapter 3
3.1 The Operating System Emulation Layer 05
Chapter 4
Buffer and Memory Management 0
4.1 Packet Buffers 06
4.2 Memory Management 08
Chapter 5
5.1 Network Interfaces 10
Chapter 6
Implementation of IP,UDP,TCP
6.1 IP Processing 12
6.2 UDP Processing 14
6.3 TCP Processing 15
Chapter 7
7.1 Interfacing the Stack 24
Chapter 8
8.1 Application Program Interface 25
SUMMARY
References
ACKNOWLEDGEMENTS
I express my sincere thanks to Prof. M.N Agnisarman Namboothiri (Head of the Department, Computer Science and Engineering, MESCE),
Mr. Zainul Abid (Staff incharge) for their kind co-operation for presenting the seminar.
I also extend my sincere thanks to all other members of the faculty of Computer Science and Engineering Department and my friends for their co-operation and encouragement.
Sahal Muhammed Haneefa
LWIP IMPLEMENTATION WITH PROXY SUPPORT
This is technique for reducing the resources needed for an implementation of the Internet protocol stack in a small device with scarce computing and memory resources . The principle of the mechanism is to move as much as possible of the resource demanding tasks from the small device to an intermediate agent known as a proxy. The proxy typically has order of magnitudes more computing and memory resources than the small device.
The proxy environment
The proxy is designed to operate in an environment as shown in Figure 3.1, where one side of the proxy is connected to the Internet through a wired link, and the other side to a wireless network with zero or more routers and possibly different wireless link technologies. Although routers may appear in the wireless network, the design of the proxy does not depend on their existence, and the proxy may be used in an environment with directly connected clients as well.
In an environment as in Figure 3.1 the wireless clients and the router, which are situated quite near each other, can communicate using a short range wireless technology such as Bluetooth. The router and the proxy communication can use a longer range and more power consuming technology, such as IEEE 802.11b.
An example of this infrastructure is the Arena project [ARN] . In this project, ice hockey players of the local hockey team will be equipped with sensors for measuring pulse rate, blood pressure, and breathing rate as well as a camera for capturing full motion video. Both the sensors and the camera will carry an implementation of the TCP/IP protocol suite, and information from the sensors and the camera will be transmitted to receivers on the Internet. The sensors, which corresponds to the wireless clients in Figure 3.1, communicates using Bluetooth technology with the camera, which is the wireless router. The camera is connected with a gateway, which runs the proxy software, using wireless LAN IEEE802.11b technology.
ABSTRACT
lwIP is an implementation of the TCP/IP protocol stack.. Interest for connecting small devices to existing network infrastructure such as global internet is steadily increasing. Such devices often has very limited CPU and memory resources and may not able to run an instance of TCP/IP protocol suite.
The focus of the lwIP stack is to reduce memory usage and code size, making lwIP suitable for use in small clients with very limited resources such as embedded systems. In order to reduce processing and memory demands, lwIP uses a tailor made API that does not require any data copying.
This report describes the design and main features that make lwIP to operate in minimal resource systems.
CHAPTER-1
1.1 INTRODUCTION
Over the last few years, the interest for connecting computers and computer supported devices to wireless networks has steadily increased. Computers are becoming more and more seamlessly integrated with everyday equipment and prices are dropping. At the same time wireless networking technologies, such as Bluetooth and IEEE 802.11b WLAN , are emerging. This gives rise to many new fascinating scenarios in areas such as health care, safety and security, transportation, and processing industry. Small devices such as sensors can be connected to an existing network infrastructure such as the global Internet, and monitored from anywhere.
The Internet technology has proven itself flexible enough to incorporate the changing network environments of the past few decades. While originally developed for low speed networks such as the ARPANET, the Internet technology today runs over a large spectrum of link technologies with vastly different characteristics in terms of bandwidth and bit error rate. It is highly advantageous to use the existing Internet technology in the wireless networks of tomorrow since a large amount of applications using the Internet technology have been developed. Also, the large connectivity of the global Internet is a strong incentive.
Since small devices such as sensors are often required to be physically small and inexpensive, an implementation of the Internet protocols will have to deal with having limited computing resources and memory. This report describes the design and implementation of a small TCP/IP stack called lwIP that is small enough to be used in minimal systems.
CHAPTER-2
OVERVIEW
As in many other TCP/IP implementations, the layered protocol design has served as a guide for the design of the implementation of lwIP. Each protocol is implemented as its own module, with a few functions acting as entry points into each protocol. Even though the protocols are implemented separately, some layer violations are made, as discussed above, in order to improve performance both in terms of processing speed and memory usage. For example, when verifying the checksum of an incoming TCP segment and when demultiplexing a segment, the source and destination IP addresses of the segment has to be known by the TCP module. Instead of passing these addresses to TCP by the means of a function call, the TCP module is aware of the structure of the IP header, and can therefore extract this information by itself.
lwIP consists of several modules. Apart from the modules implementing the TCP/IP protocols (IP, ICMP, UDP, and TCP) a number of support modules are implemented.
The support modules consists of :-
¢ The operating system emulation layer (described in Chapter3)
¢ The buffer and memory management subsystems
(described in Chapter 4)
¢ Network interface functions (described in Chapter 5)
¢ Functions for computing Internet checksum (Chapter 6)
¢ An abstract API (described in Chapter 8 )
2.1 PROTOCOL LAYERING
The protocols in the TCP/IP suite are designed in a layered fashion, where each protocol layer solves a separate part of the communication problem. This layering can serve as a guide for designing the implementation of the protocols, in that each protocol can be implemented separately from the other. Implementing the protocols in a strictly layered way can however, lead to a situation where the communication overhead between the protocol layers degrades the overall performance. To overcome these problems, certain internal aspects of a protocol can be made known to other protocols. Care must be taken so that only the important information is shared among the layers.
Most TCP/IP implementations keep a strict division between the application layer and the lower protocol layers, whereas the lower layers can be more or less interleaved. In most operating systems, the lower layer protocols are implemented as a part of the operating system kernel with entry points for communication with the application layer process. The application program is presented with an abstract view of the TCP/IP implementation, where network communication differs only very little from inter-process communication or file I/O. The implications of this is that since the application program is unaware of the buffer mechanisms used by the lower layers, it cannot utilize this information to, e.g., reuse buffers with frequently used data. Also, when the application sends data, this data has to be copied from the application process' memory space into internal buffers before being processed by the network code.
The operating systems used in minimal systems such as the target system of lwIP most often do not maintain a strict protection barrier between the kernel and the application processes. This allows using a more relaxed scheme for communication between the application and the lower layer protocols by the means of shared memory. In particular, the application layer can be made aware of the buffer handling mechanisms used by the lower layers. Therefore, the application can more efficiently reuse buffers. Also, since the application process can use the same memory as the networking code the application can read and write directly to the internal buffers, thus saving the expense of performing a copy.
2.2 PROCESS MODEL
The process model of a protocol implementation describes in which way the system has been divided into different processes. One process model that has been used to implement communication protocols is to let each protocol run as a stand alone process. With this model, a strict protocol layering is enforced, and the communication points between the protocols must be strictly defined.
While this approach has its advantages such as protocols can be added at runtime, understanding the code and debugging is generally easier, there are also disadvantages. The strict layering is not, as described earlier, always the best way to implement protocols. Also, and more important, for each layer crossed, a context switch must be made. For an incoming TCP segment this would mean three context switches, from the device driver for the network interface, to the IP process, to the TCP process and finally to the application process. In most operating systems a context switch is fairly expensive.
Another common approach is to let the communication protocols reside in the kernel of the operating system. In the case of a kernel implementation of the communication protocols, the application processes communicate with the protocols through system calls. The communication protocols are not strictly divided from each other but may use the techniques of crossing the protocol layering.
lwIP uses a process model in which all protocols reside in a single process and are thus separated from the operating system kernel. Application programs may either reside in the lwIP process, or be in separate processes. Communication between the TCP/IP stack and the application programs are done either by function calls for the case where the application program shares a process with lwIP, or by the means of a more abstract API.
Having lwIP implemented as a user space process rather than in the operating system kernel has both its advantages and disadvantages. The main advantage of having lwIP as a process is that is portable across different operating systems. Since lwIP is designed to run in small operating systems that generally do not support neither swapping out processes not virtual memory, the delay caused by having to wait for disk activity if part of the lwIP process is swapped or paged out to disk will not be a problem. The problem of having to wait for a scheduling quantum before getting a chance to service requests still is a problem however, but there is nothing in the design of lwIP that precludes it from later being implemented in an operating system kernel.
CHAPTER-3
THE OPERATING SYSTEM EMULATION LAYER
In order to make lwIP portable, operating system specific function calls and data structures are not used directly in the code. Instead, when such functions are needed the operating system emulation layer is used. The operating system emulation layer provides a uniform interface to operating system services such as timers, process synchronization, and message passing mechanisms. In principle, when porting lwIP to other operating systems only an implementation of the operating system emulation layer for that particular operating system is needed.
The operating system emulation layer provides a timer functionality that is used by TCP. The timers provided by the operating system emulation layer are one-shot timers with a granularity of at least 200 ms that calls a registered function when the time-out occurs.
The only process synchronization mechanism provided is semaphores. Even if semaphores are not available in the underlying operating system they can be emulated by other synchronization primitives such as conditional variables or locks.
The message passing is done through a simple mechanism which uses an abstraction called mailboxes. A mailbox has two operations: post and fetch. The post operation will not block the process; rather, messages posted to a mailbox are queued by the operating system emulation layer until another process fetches them. Even if the underlying operating system does not have native support for the mailbox mechanism, they are easily implemented using semaphores.
CHAPTER-4
BUFFER AND MEMORY MANAGEMENT
The memory and buffer management system in a communication system must be prepared to accommodate buffers of very varying sizes, ranging from buffers containing full-sized TCP segments with several hundred bytes worth of data to short ICMP echo replies consisting of only a few bytes. Also, in order to avoid copying it should be possible to let the data content of the buffers reside in memory that is not managed by the networking subsystem, such as application memory or ROM.
4.1 Packet buffers | pbufs
A pbuf is lwIP's internal representation of a packet, and is designed for the special needs of the minimal stack. Pbufs are similar to the mbufs used in the BSD implementations. The pbuf structure has support both for allocating dynamic memory to hold packet contents, and for letting packet data reside in static memory. Pbufs can be linked together in a list, called a pbuf chain so that a packet may span over several pbufs.
Pbufs are of three types, PBUF RAM, PBUF ROM, and PBUF POOL. The pbuf shown in Figure 1 represents the PBUF RAM type, and has the packet data stored in memory managed by the pbuf subsystem. The pbuf in Figure 2 is an example of a chained pbuf, where the first pbuf in the chain is of the PBUF RAM type, and the second is of the PBUF ROM type, which means that it has the data located in memory not managed by the pbuf system. The third type of pbuf, PBUF POOL, is shown in Figure 3 and consists of fixed size pbufs allocated from a pool of fixed size pbufs. A pbuf chain may consist of multiple types of pbufs.
The three types have different uses. Pbufs of type PBUF POOL are mainly used by network device drivers since the operation of allocating a single pbuf is fast and is therefore suitable for use in an interrupt handler. PBUF ROM pbufs are used when an application sends data that is located in memory managed by the application. This data may not be modified after the pbuf has been handed over to the TCP/IP stack and therefore this pbuf type main use is when the data is located in ROM (hence the name PBUF ROM). Headers that are prepended to the data in a PBUF ROM pbuf are stored in a PBUF RAM pbuf that is chained to the front of of the PBUF ROM pbuf, as in Figure 2
Pbufs of the PBUF RAM type are also used when an application sends data that is dynamically generated. In this case, the pbuf system allocates memory not only for the application data, but also for the headers that will be prepended to the data. This is seen in Figure 1. The pbuf system cannot know in advance what headers will be prepended to the data and assumes the worst case.The size of the headers is configurable at compile time.
In essence, incoming pbufs are of type PBUF POOL and outgoing pbufs are of the PBUF ROM or PBUF RAM types.
The internal structure of a pbuf can be seen in the Figures 1 through 3. The pbuf structure consists of two pointers, two length fields, flags field, and a reference count. The next field is a pointer to the next pbuf in case of a pbuf chain. The payload pointer points to the start of the data in the pbuf. The len field contains the length of the data contents of the pbuf. The tot len field contains the sum of the length of the current pbuf and all len fields of following pbufs in the pbuf chain. In other words, the tot len field is the sum of the len field and the value of the tot len field in the following pbuf in the pbuf chain. The flags field indicates the type of the pbuf and the ref field -contains a reference count. The next and payload fields are native pointers and the size of those varies depending on the processor architecture used. The two length fields are 16 bit unsigned integers and the flags and ref fields are 4 bit wide. The total size of the pbuf structure depends on the size of a pointer in the processor architecture being used and on the smallest alignment possible for the processor architecture. On an architecture with 32 bit pointers and 4 byte alignment, the total size is 16 bytes and on an architecture with 16 bit pointers and 1 byte alignment, the size is 9 bytes.
4.2 Memory management
The memory manager supporting the pbuf scheme is very simple. It handles allocations and deallocations of contiguous regions of memory and can shrink the size of a previously allocated memory block. The memory manager uses a dedicated portion of the total memory in the system. This ensures that the networking system does not use all of the available memory, and that the operation of other programs is not disturbed if the networking system has used all of it's memory.
Internally, the memory manager keeps track of the allocated memory by placing a small structure on top of each allocated memory block. This structure holds two pointers to the next and previous allocation block in memory. It also has a used flag which indicates whether the allocation block is allocated or not.
Memory is allocated by searching the memory for an unused allocation block that is large enough for the requested allocation. The first-fit principle is used so that the first block that is large enough is used. When an allocation block is deallocated, the used flag is set to zero. In order to prevent fragmentation, the used flag of the next and previous allocation blocks are checked. If any of them are unused, the blocks are combined into one larger unused block.
CHAPTER-5
NETWORK INTERFACES
In lwIP device drivers for physical network hardware are represented by a network interface structure. The network interfaces are kept on a global linked list, which is linked by the next pointer in the structure. Each network interface has a name, stored in the name field . This two letter name identifies the kind of device driver used for the network interface and is only used when the interface is configured by a human operator at runtime. The name is set by the device driver and should reflect the kind of hardware that is represented by the network interface. For example, a network interface for a Bluetooth driver might have the name bt and a network interface for IEEE 802.11b WLAN hardware could have the name wl. Since the names not necessarily are unique, the num field is used to distinguish different network interfaces of the same kind.
Figure 4. The memory allocation structure.
The three IP addresses ip addr, netmask and gw are used by the IP layer when sending and receiving packets, and their use is described in the next section. It is not possible to configure a network interface with more than one IP address. Rather, one network interface would have to be created for each IP address.
The input pointer points to the function the device driver should call when a packet has been received.A network interface is connected to a device driver through the output pointer. This pointer points to a function in the device driver that transmits a packet on the physical network and it is called by the IP layer when a packet is to be sent. This field is filled by the initialization
function of the device driver. The third argument to the output function, ipaddr, is the IP address of the host that should receive the actual link layer frame. It does not have to be the same as the destination address of the IP packet. In particular, when sending an IP packet to a host that is not on the local network, the link level frame will be sent to a router on the network. In this case, the IP address given to the output function will be the IP address of the router.
Finally, the state pointer points to device driver specific state for the network interface and is set by the device driver.
CHAPTER-6
IMPLEMENTATION OF IP,UDP,TCP
6.1 IP Processing
lwIP implements only the most basic functionality of IP. It can send, receive and forward packets, but cannot send or receive fragmented IP packets nor handle packets with IP options. For most applications this does not pose any problems.
6.1.1 Receiving packets
For incoming IP packets, processing begins when the ip input() function is called by a network device driver. Here, the initial sanity checking of the IP version field and the header length is done, as well as computing and checking the header checksum. It is expected that the stack will not receive any IP fragments since the proxy is assumed to reassemble any fragmented packets, thus any packet that is an IP fragment is silently discarded. Packets carrying IP options are also assumed to be handled by the proxy, and are dropped.
Next, the function checks the destination address with the IP addresses of the network interfaces to determine if the packet was destined for the host. The network interfaces are ordered in a linked list, and it is searched linearly. The number of network interfaces is expected to be small so a more sophisticated search strategy than a linear search has not been implemented.
If the incoming packet is found to be destined for this host, the protocol field is used to decide to which higher level protocol the packet should be passed to.
6.1.2 Sending packets
An outgoing packet is handled by the function ip output(), which uses the function ip route() to find the appropriate network interface to transmit the packet on. When the outgoing network interface is determined, the packet is passed to ip output if() which takes the outgoing network interface as an argument. Here, all IP header fields are filled in and the IP header checksum is computed. The source and destination addresses of the IP packet is passed as an argument to ip output if(). The source address may be left out, however, and in this case the IP address of the outgoing network interface is used as the source IP address of the packet.
The ip route() function finds the appropriate network interface by linearly searching the list of network interfaces. During the search the destination IP address of the IP packet is masked with the netmask of the network interface. If the masked destination address is equal to the masked IP address of the interface, the interface is chosen. If no match is found, a default network interface is used. The default network interface is configured either at boot-up or at runtime by a human operator. If the network address of the default interface does not match the destination IP address, the gw field in the network interface structure is chosen as the destination IP address of the link level frame. (Notice that the destination address of the IP packet and the destination address of the link level frame will be different in this case.) This primitive form of routing glosses over the fact that a network might have many routers attached to it. For the most basic case, where a local network only has one router, this works however.
Since the transport layer protocols UDP and TCP need to have the destination IP address when computing the transport layer checksum, the outgoing network interface must in some cases be determined before the packet is passed to the IP layer. This is done by letting the transport layer functions call the ip route() function directly, and since the outgoing network interface is known already when the packet reaches the IP layer, there is no need to search the network interface list again. Instead, those protocols call the ip output if() function directly. Since this function takes a network interface as an argument, the search for an outgoing interface is avoided.
Human configuration of lwIP during runtime requires an application program that is able to configure the stack. Such a program is not included in lwIP.
6.1.3 Forwarding packets
If none of the network interfaces has the same IP address as an incoming packet's destination address, the packet should be forwarded. This is done by the function ip forward(). Here, the TTL field is decreased and if it reaches zero, an ICMP error message is sent to the original sender of the IP packet and the packet is discarded. Since the IP header is changed, the IP header checksum needs to be adjusted. The is no need to recompute the entire checksum, however, since simple arithmetic can be used to adjust the original IP checksum.
Finally, the packet is forwarded to the appropriate network interface. The algorithm used to find the appropriate network interface is the same that is used when sending IP packets.
6.1.4 ICMP processing
ICMP processing is fairly simple. ICMP packets received by ip input() are handed over to icmp input(), which decodes the ICMP header and takes the appropriate action. Some ICMP messages are passed to upper layer protocols and those are taken care of by special functions in the transport layer. ICMP destination unreachable messages can be sent by transport layer protocols, in particular by UDP, and the function icmp dest unreach() is used for this.
Using ICMP ECHO messages to probe a network is widely used, and therefore ICMP echo processing is optimized for performance. The actual processing takes place in icmp input(), and consists of swapping the IP destination and source addresses of the incoming packet, change the ICMP type to echo reply and adjust the ICMP checksum. The packet is then passed back to the IP layer for transmission.
6.2 UDP processing
UDP is a simple protocol used for demultiplexing packets between different processes. The state for each UDP session is kept in a PCB structure . The UDP PCBs are kept on a linked list which is searched for a match when a UDP datagram arrives.
The UDP PCB structure contains a pointer to the next PCB in the global linked list of UDP PCBs. A UDP session is defined by the IP addresses and port numbers of the end-points and these are stored in the local ip, dest ip, local port and dest port fields. The flags field indicates what UDP checksum policy that should be used for this session. This can be either to switch UDP checksumming off completely, or to use UDP Lite in which the checksum covers only parts of the datagram. If UDP Lite is used, the chksum len field specifies how much of the datagram that should be checksummed.
The last two arguments recv and recv arg are used when a datagram is received in the session specified by the PCB. The function pointed to by recv is called when a datagram is received.
Due to the simplicity of UDP, the input and output processing is equally simple and follows a fairly straight line (Figure 7 ). To send data, the application program calls udp send() which calls upon udp output(). Here the the necessary checksumming is done and UDP header fields are filled. Since the checksum includes the IP source address of the IP packet, the function ip route() is in some cases called to find the network interface to which the packet is to be transmitted. The IP address of this network interface is used as the source IP address of the packet. Finally, the packet is turned over to ip output if() for transmission.
fig 7 UDP processing
When a UDP datagram arrives, the IP layer calls the udp input() function. Here, if check-summing should be used in the session, the UDP checksum is checked and the datagram is demultiplexed. When the corresponding UDP PCB is found, the recv function is called.
6.3 TCP processing
TCP is a transport layer protocol that provides a reliable byte stream service to the application layer. TCP is more complex than the other protocols described here, and the TCP code constitutes50% of the total code size of lwIP.
6.3.1 Overview
The basic TCP processing (Figure 8 ) is divided into six functions; the functions tcp input(),tcp process(), and tcp receive() which are related to TCP input processing, and tcp write(),tcp enqueue(), and tcp output() which deals with output processing.
Fig 8 TCP processing
When an application wants to send TCP data, tcp write() is called. The function tcp write() passes control to tcp enqueue() which will break the data into appropriate sized TCP segmentsif necessary and put the segments on the transmission queue for the connection. The function tcp output() will then check if it is possible to send the data, i.e., if there is enough space in the receiver's window and if the congestion window is large enough and if so, sends the data using ip route() and ip output if().
Input processing begins when ip input() after verifying the IP header hands over a TCP segment to tcp input(). In this function the initial sanity checks (i.e., checksumming and TCP options parsing) are done as well as deciding to which TCP connection the segment belongs. The segment is then processed by tcp process(), which implements the TCP state machine, and any necessary state transitions are made. The function tcp receive() will be called if the connection is in a state to accept data from the network. If so, tcp receive() will pass the segment up to an application program. If the segment constitutes an ACK for unacknowledged (thus previously buffered) data, the data is removed from the buffers and its memory is reclaimed. Also, if an ACK for data was received the receiver might be willing to accept more data and therefore tcp output() is called.
6.3.2 Data structures
The data structures used in the implementation of TCP are kept small due to the memory constraints in the minimal system for which lwIP is intended. There is a tradeoff between the complexity of the data structures and the complexity of the code that uses the data structures, and in this case the code complexity has been sacrificed in order to keep the size of the data structures small.
The TCP PCB is fairly large. Since TCP connections in the LISTEN and TIME-WAIT needs to keep less state information than connections in other states, a smaller PCB data structure is used for those connections. This data structure is overlaid with the full PCB structure, and the ordering of the items in the PCB structure is therefore somewhat awkward.
The TCP PCBs are kept on a linked list, and the next pointer links the PCB list together. The state variable contains the current TCP state of the connection. Next, the IP addresses and port numbers which identify the connection are stored. The mss variable contains the maximum segment size allowed for the connection.
The rcv nxt and rcv wnd fields are used when receiving data. The rcv nxt field contains the next sequence number expected from the remote end and is thus used when sending ACKs to the remote host. The receiver's window is kept in rcv wnd and this is advertised in outgoing TCP segments. The field tmr is used as a timer for connections that should be removed after a certain amount of time, such as connections in the TIME-WAIT state. The maximum segment size allowed on the connection is stored in the mss field. The flags field contains additional state information of the connection, such as whether the connection is in fast recovery or if a delayed ACK should be send.
The fields rttest, rtseq, sa, and sv are used for the round-trip time estimation. The sequence number of the segment that is used for estimating the round-trip time is stored in rtseq and the time this segment was sent is stored in rttest. The average round-trip time and the round-trip time variance is stored in sa and sv. These variables are used when calculating the retransmission time-out which is stored in the rto field.
The two fields lastack and dupacks are used in the implementation of fast retransmit and fast recovery. The lastack field contains the sequence number acknowledged by the last ACK received and dupacks contains a count of how many ACKs that has been received for the sequence number in lastack. The current congestion window for the connection is stored in the cwnd field and the slow start threshold is kept in ssthresh.
The six fields snd ack, snd nxt, snd wnd, snd wl1, snd wl2 and snd lbb are used when sending data. The highest sequence number acknowledged by the receiver is stored in snd ack and the next sequence number to send is kept in snd nxt. The receiver's advertised window is held in snd wnd and the two fields snd wl1 and snd wl2 are used when updating snd wnd. The snd lbb field contains the sequence number of the last byte queued for transmission.
The function pointer recv and recv arg are used when passing received data to the application layer. The three queues unsent, unacked and ooseq are used when sending and receiving data. Data that has been received from the application but has not been sent is queued in unsent and data that has been sent but not yet acknowledged by the remote host is held in unacked. Received data that is out of sequence is buffered in ooseq.
struct tcp_seg {
struct tcp_seg *next;
u16_t len;
struct pbuf *p;
struct tcp_hdr *tcphdr;
void *data;
u16_t rtime;
};
Figure 9. The tcp seg structure
The tcp seg structure in Figure 9 is the internal representation of a TCP segment. This structure starts with a next pointer which is used for linking when queuing segments. The len field contains the length of the segment in TCP terms. This means that the len field for a data segment will contain the length of the data in the segment, and the len field for an empty segment with the SYN or FIN flags set will be 1. The pbuf p is the buffer containing the actual segment and the tcphdr and data pointers points to the TCP header and the data in the segment, respectively.
For outgoing segments, the rtime field is used for the retransmission time-out of this segment. Since incoming segments will not need to be retransmitted, this field is not needed and memory for this field is not allocated for incoming segments.
6.3.3 Sequence number calculations
The TCP sequence numbers that are used to enumerate the bytes in the TCP byte stream are unsigned 32 bit quantities, hence in the range [0; 232 ¡ 1]. Since the number of bytes sent in a TCP connection might be more than the number of 32-bit combinations, the sequence numbers are calculated modulo 232. This means that ordinary comparison operators cannot be used withTCP sequence numbers. The modified comparison operators, called <seq and >seq, are defined by the relations
s <seq t , s ¡ t < 0
and
s >seq t , s ¡ t > 0;
where s and t are TCP sequence numbers. The comparison operators for ¢ and ¸ are defined equivalently. The comparison operators are defined as C macros in the header file.
6.3.4 Queuing and transmitting data
Data that is to be sent is divided into appropriate sized chunks and given sequence numbers by the tcp enqueue() function. Here, the data is packeted into pbufs and enclosed in a tcp seg struct The TCP header is build in the pbuf, and filled in with all fields except the acknowledgment number, ackno, and the advertised window, wnd. These fields can change during the queuing time of the segment and are therefore set by tcp output() which does the actual transmission of the segment. After the segments are built, they are queued on the unsent list in the PCB. The tcp enqueue() function tries to fill each segment with a maximum segment size worth of data and when an under-full segment is found at the end of the unsent queue, this segment is appended with the new data using the pbuf chaining functionality.
After tcp enqueue() has formatted and queued the segments, tcp output() is called. It checks if there is any room in the current window for any more data. The current window is computed by taking the maximum of the congestion window and the advertised receiver's window. Next, it fills in the fields of the TCP header that was not filled in by tcp enqueue() and transmits the segment using ip route() and ip output if(). After transmission the segment is put on the unacked list, on which it stays until an ACK for the segment has been received.
When a segment is on the unacked list, it is also timed for retransmission . When a segment is retransmitted the TCP and IP headers of the original segment is kept and only very little changes has to be made to the TCP header. The ackno and wnd fields of the TCP header are set to the current values since we could have received data during the time between the original transmission of the segment and the retransmission. This changes only two 16-bit words in the header and the whole TCP checksum does not have to be recomputed since simple arithmetic can be used to update the checksum. The IP layer has already added the IP header when the segment was originally transmitted and there is no reason to change it. Thus a retransmission does not require any recomputation of the IP header checksum.
6.3.4.1 Silly window avoidance
The Silly Window Syndrome (SWS) is a TCP phenomena that can lead to very bad performance. SWS occurs when a TCP receiver advertises a small window and the TCP sender immediately sends data to fill the window. When this small segment is acknowledged the window is opened again by a small amount and sender will again send a small segment to fill the window. This leads to a situation where the TCP stream consists of very small segments. In order to avoid SWS both the sender and the receiver must try to avoid this situation. The receiver must not advertise small window updates and the sender must not send small segments when only a small window is offered.
In lwIP SWS is naturally avoided at the sender since TCP segments are constructed and queued without knowledge of the advertised receiver's window. In a large transfer the output queue will consist of maximum sized segments. This means that if a TCP receiver advertises a small window, the sender will not send the first segment on the queue since it is larger than the advertised window. Instead, it will wait until the window is large enough for a maximum sized segment.
When acting as a TCP receiver, lwIP will not advertise a receiver's window that is smaller than the maximum segment size of the connection.
6.3.5 Receiving segments
6.3.5.1 Demultiplexing
When TCP segments arrive at the tcp input() function, they are demultiplexed between the TCP PCBs. The demultiplexing key is the source and destination IP addresses and the TCP port numbers. There are two types of PCBs that must be distinguished when demultiplexing a segment; those that correspond to open connections and those that correspond to connections that are half open. Half open connections are those that are in the LISTEN state and only have the local TCP port number specified and optionally the local IP address, whereas open connections have the both IP addresses and both port numbers specified.
Many TCP implementations, use a technique where a linked list of PCBs with a single entry cache is used. The rationale behind this is that most TCP connections constitute bulk transfers which typically show a large amount locality, resulting in a high cache hit ratio. Other caching schemes include keeping two one entry caches, one for the PCB corresponding to the last packet that was sent and one for the PCB of the last packet received . An alternative scheme to exploit locality can be done by moving the most recently used PCB to the front of the list. Both methods have been shown to outperform the one entry cache scheme.
In lwIP, whenever a PCB match is found when demultiplexing a segment, the PCB is moved to the front of the list of PCBs. PCBs for connections in the LISTEN state are not moved to the front however, since such connections are not expected to receive segments as often as connections that are in a state in which they receive data.
6.3.5.2 Receiving data
The actual processing of incoming segments is made in the function tcp receive(). The acknowledgment number of the segment is compared with the segments on the unacked queue of the connection. If the acknowledgment number is higher than the sequence number of a segment on the unacked queue, that segment is removed from the queue and the allocated memory for the segment is deallocated.
An incoming segment is out of sequence if the sequence number of the segment is higher than the rcv nxt variable in the PCB. Out of sequence segments are queued on the ooseq queue in the PCB. If the sequence number of the incoming segment is equal to rcv nxt, the segment is delivered to the upper layer by calling the recv function in the PCB and rcv nxt is increased by the length of the incoming segment. Since the reception of an in-sequence segment might mean that a previously received out of sequence segment now is the next segment expected, the ooseq queued is checked. If it contains a segment with sequence number equal to rcv nxt, this segment is delivered to the application by a call to to recv function and rcv nxt is updated. This process continues until either the ooseq queue is empty or the next segment on ooseq is out of sequence.
6.3.6 Accepting new connections
TCP connections that are in the LISTEN state, i.e., that are passively open, are ready to accept new connections from a remote host. For those connections a new TCP PCB is created and must be passed to the application program that opened the initial listening TCP connection. In lwIP this is done by letting the application register a callback function that is to be called when a new connection has been established.
When a connection in the LISTEN state receives a TCP segment with the SYN flag set, a new connection is created and a segment with the SYN and ACK flags are sent in response to the SYN segment. The connection then enters the SYN-RCVD state and waits for an acknowledgment for the sent SYN segment. When the acknowledgment arrives, the connection enters the ESTABLISHED state, and the accept function (the accept field in the PCB structure is called.
6.3.7 Fast retransmit
Fast retransmit and fast recovery is implemented in lwIP by keeping track of the last sequence number acknowledged. If another acknowledgment for the same sequence number is received, the dupacks counter in the TCP PCB is increased. When dupacks reaches three, the first segment on the unacked queue is retransmitted and fast recovery is initialized. The implementation of fast recovery follows the steps laid out in . Whenever an ACK for new data is received, the dupacks counter is reset to zero.
6.3.8 Timers
As in the the BSD TCP implementation, lwIP uses two periodical timers that goes off every 200 ms and 500 ms. Those two timers are then used to implement more complex logical timers such as the retransmission timers, the TIME-WAIT timer and the delayed ACK timer.
The fine grained timer, tcp timer fine() goes through every TCP PCB checking if there are any delayed ACKs that should be sent, as indicated by the flag field in the tcp pcb structure . If the delayed ACK flag is set, an empty TCP acknowledgment segment is sent and the flag is cleared.
The coarse grained timer, implemented in tcp timer coarse(), also scans the PCB list. For every PCB, the list of unacknowledged segments (, is traversed, and the rtime variable is increased. If rtime becomes larger than the current retransmission time-out as given by the rto variable in the PCB, the segment is retransmitted and the retransmission time-out is doubled. A segment is retransmitted only if allowed by the values of the congestion window and the advertised receiver's window. After retransmission, the congestion window is set to one maximum segment size, the slow start threshold is set to half of the effective window size, and slow start is initiated on the connection.
For connections that are in TIME-WAIT, the coarse grained timer also increases the tmr field in the PCB structure. When this timer reaches the 2*MSL threshold, the connection is removed.The coarse grained timer also increases a global TCP clock, tcp ticks. This clock is used for round-trip time estimation and retransmission time-outs.
6.3.9 Round-trip time estimation
The round-trip time estimation is a critical part of TCP since the estimated round-trip time is used when determining a suitable retransmission time-out. In lwIP round-trip times measurements are taken in a fashion similar to the BSD implementations. Round-trip times are measured once per round-trip and the smoothing function described in is used for the calculation of a suitable retransmission time-out.
The TCP PCB variable rtseq hold the sequence number of the segment for which the roundtrip time is measured. The rttest variable in the PCB holds the value of tcp ticks when the segment was first transmitted. When an ACK for a sequence number equal to or larger than rtseq is received, the round-trip time is measured by subtracting rttest from tcp ticks. If a retransmission occurred during the round-trip time measurement, no measurement is taken.
6.3.10 Congestion control
The implementation of congestion control is surprisingly simple and consists of a few lines of code in the output and input code. When an ACK for new data is received the congestion window, cwnd, is increased either by one maximum segment size or by mss2/cwnd, depending on whether the connection is in slow start or congestion avoidance. When sending data the minimum value of the receiver's advertised window and the congestion window is used to determine how much data that can be sent in each window.
CHAPTER-7
INTERFACING THE STACK
There are two ways for using the services provided by the TCP/IP stack; either by calling the functions in the TCP and UDP modules directly, or to use the lwIP API presented in the next section.
The TCP and UDP modules provide a rudimentary interface to the networking services. The interface is based on callbacks and an application program that uses the interface can therefore not operate in a sequential manner. This makes the application program harder to program and the application code is harder to understand. In order to receive data the application program registers a callback function with the stack. The callback function is associated with a particular connection and when a packet arrives in the connection, the callback function is called by the stack.
Furthermore, an application program that interfaces the TCP and UDP modules directly has to (at least partially) reside in the same process as the TCP/IP stack. This is due to the fact that a callback function cannot be called across a process boundary. This has both advantages and disadvantages. One advantage is that since the application program and the TCP/IP stack are in the same process, no context switching will be made when sending or receiving packets. The main disadvantage is that the application program cannot involve itself in any long running computations since TCP/IP processing cannot occur in parallel with the computation, thus degrading communication performance. This can be overcome by splitting the application into two parts, one part dealing with the communication and one part dealing with the computation. The part doing the communication would then reside in the TCP/IP process and the computationally heavy part would be a separate process. The lwIP API presented in the next section provides a structured way to divide the application in such a way.
CHAPTER-8
APPLICATION PROGRAM INTERFACE
Due to the high level of abstraction provided by the BSD socket API, it is unsuitable for use in a minimal TCP/IPimplementation. In particular, BSD sockets require data that is to be sent to be copied from the application program to internal buffers in the TCP/IP stack. The reason for copying the data is that the application and the TCP/IP stack usually reside in different protection domains. In most cases the application program is a user process and the TCP/IP stack resides in the operating system kernel. By avoiding the extra copy, the performance of the API can be greatly improved . Also, in order to make a copy, extra memory needs to be allocated for the copy, effectively doubling the amount of memory used per packet.
The lwIP API was designed for lwIP and utilizes knowledge of the internal structure of lwIP to achieve effectiveness. The lwIP API is very similar to the BSD API, but operates at a slightly lower level. The API does not require that data is copied between the application program and the TCP/IP stack, since the application program can manipulate the internal buffers directly.
.
8.1 BASIC CONCEPTS
From the application's point of view, data handling in the BSD socket API is done in continuous memory regions. This is convenient for the application programmer since manipulation of data in application programs is usually done in such continuous memory chunks. Using this type of mechanism with lwIP would not be advantageous, since lwIP usually handles data in buffers where the data is partitioned into smaller chunks of memory. Thus the data would have to be copied into a continuous memory area before being passed to the application. This would waste both processing time and memory since. Therefore, the lwIP API allows the application program to manipulate data directly in the partitioned buffers in order to avoid the extra copy.
The lwIP API uses a connection abstraction similar to that of the BSD socket API. There are very noticeable differences however; where an application program using the BSD socket API need not be aware of the distinction between an ordinary file and a network connection, an application program using the lwIP API has to be aware of the fact that it is using a network connection. Network data is received in the form of buffers where the data is partitioned into smaller chunks of memory. Since many applications wants to manipulate data in a continuous memory region, a convenience function for copying the data from a fragmented buffer to continuous memory exists.
Sending data is done differently depending on whether the data should be sent over a TCP connection or as UDP datagrams. For TCP, data is sent by passing the output function a pointer to a continuous memory region. The TCP/IP stack will partition the data into appropriately sized packets and queue them for transmission. When sending UDP datagrams, the application program will to explicitly allocate a buffer and fill it with data. The TCP/IP stack will send the datagram immediately when the output function is called.
8.2 IMPLEMENTATION OF THE API
The implementation of the API is divided into two parts, due to the process model of the TCP/IP stack. As shown in Figure 10 , parts of the API is implemented as a library linked to the application program, and parts are implemented in the TCP/IP process. The two parts communicate using the interprocess communication (IPC) mechanisms provided by the operating system emulation layer. The current implementation uses the following three IPC mechanisms:
¢ shared memory
¢ message passing
¢ semaphores.
While these IPC types are supported by the operating system layer, they need not be directly supported by the underlying operating system. For operating systems that do not natively support them, the operating system emulation layer emulates them.
Fig 10 Division of API implementation
The general design principle used is to let as much work as possible be done within the application process rather than in the TCP/IP process. This is important since all processes use the TCP/IP process for their TCP/IP communication. Keeping down the code footprint of the part of the API that is linked with the applications is not as important. This code can be shared among the processes, and even if shared libraries are not supported by the operating system, the code is stored in ROM. Embedded systems usually carry fairly large amounts of ROM, whereas processing power is scarce.
The buffer management is located in the library part of the API implementation. Buffers are created, copied and deallocated in the application process. Shared memory is used to pass the buffers between the application process and the TCP/IP process. The buffer data type used in communication with the application program is an abstraction of the pbuf data type.
Buffers carrying referenced memory, as opposed to allocated memory, is also passed using shared memory. For this to work, is has to be possible to share the referenced memory between the processes. The operating systems used in embedded systems for which lwIP is intended usually do not implement any form of memory protection, so this will not be a problem.
The functions that handle network connections are implemented in the part of the API implementation that resides in the TCP/IP process. The API functions in the part of the API that runs in the application process will pass a message using a simple communication protocol to the API implementation in the TCP/IP process. The message includes the type of operation that should be carried out and any arguments for the operation. The operation is carried out by the API implementation in the TCP/IP process and the return value is sent to the application process by message passing.
8.3 PERFORMANCE ANALYSIS
The performance of lwIP in terms of RAM usage or code effciency have not been formally tested and this has been noted in future work. Simple tests have been conducted, however, and these have shown that lwIP running a simple HTTP/1.0 server is able to serve at least 10 simultaneous requests for web pages while consuming less than 4 kilobytes of RAM. In those tests,only the memory used by the protocols, buffering system, and the application program has been taken into account.
SUMMARY
The small TCP/IP stack
lwIP, the small TCP/IP stack, has been shown to have a small code size. It is designed for systems with little RAM and simple tests have shown that it can operate in an environment with very little RAM available. Operating system dependencies are moved into a separate module to simplify porting of lwIP to other operating systems. The TCP/IP stack has support for TCP, UDP, ICMP, and IP. The TCP layer is fully functional and implements congestion control, round-trip time estimation, fast retransmit, and fast recovery. The IP layer has support for IP forwarding and can be used as a simple router.
REFERENCES
¢ Adam Dunkels, lightweight IP: An overview.Swedish Institute of Computer Science
¢ B. Ahlgren, M. BjÄorkman, and K. Moldeklev. The performance of a no-copy api for communication subsystems.
¢ M. Allman, V. Paxson, and W. Stevens. TCP congestion control. RFC 2581, Internet Engineering Task Force.
¢ D.D.Clark. Modularity and efficiency in protocol implementation. RFC 817, Internet Engineering Task Force.
¢ Paul E. McKenney and Ken F. Dove. Efficient demultiplexing of incoming TCP packets.
¢ T. Mallory and A. Kullberg. Incremental updating of the internet checksum. RFC 1141, Internet Engineering Task Force.
CONTENTS
Chapter 1
1.1 Introduction 01
Chapter 2
Overview 0
2.1 Protocol Layering 02
2.2 Process Model 03
Chapter 3
3.1 The Operating System Emulation Layer 05
Chapter 4
Buffer and Memory Management 0
4.1 Packet Buffers 06
4.2 Memory Management 08
Chapter 5
5.1 Network Interfaces 10
Chapter 6
Implementation of IP,UDP,TCP
6.1 IP Processing 12
6.2 UDP Processing 14
6.3 TCP Processing 15
Chapter 7
7.1 Interfacing the Stack 24
Chapter 8
8.1 Application Program Interface 25
SUMMARY
References
ACKNOWLEDGEMENTS
I express my sincere thanks to Prof. M.N Agnisarman Namboothiri (Head of the Department, Computer Science and Engineering, MESCE),
Mr. Zainul Abid (Staff incharge) for their kind co-operation for presenting the seminar.
I also extend my sincere thanks to all other members of the faculty of Computer Science and Engineering Department and my friends for their co-operation and encouragement.
Sahal Muhammed Haneefa
LWIP IMPLEMENTATION WITH PROXY SUPPORT
This is technique for reducing the resources needed for an implementation of the Internet protocol stack in a small device with scarce computing and memory resources . The principle of the mechanism is to move as much as possible of the resource demanding tasks from the small device to an intermediate agent known as a proxy. The proxy typically has order of magnitudes more computing and memory resources than the small device.
The proxy environment
The proxy is designed to operate in an environment as shown in Figure 3.1, where one side of the proxy is connected to the Internet through a wired link, and the other side to a wireless network with zero or more routers and possibly different wireless link technologies. Although routers may appear in the wireless network, the design of the proxy does not depend on their existence, and the proxy may be used in an environment with directly connected clients as well.
In an environment as in Figure 3.1 the wireless clients and the router, which are situated quite near each other, can communicate using a short range wireless technology such as Bluetooth. The router and the proxy communication can use a longer range and more power consuming technology, such as IEEE 802.11b.
An example of this infrastructure is the Arena project [ARN] . In this project, ice hockey players of the local hockey team will be equipped with sensors for measuring pulse rate, blood pressure, and breathing rate as well as a camera for capturing full motion video. Both the sensors and the camera will carry an implementation of the TCP/IP protocol suite, and information from the sensors and the camera will be transmitted to receivers on the Internet. The sensors, which corresponds to the wireless clients in Figure 3.1, communicates using Bluetooth technology with the camera, which is the wireless router. The camera is connected with a gateway, which runs the proxy software, using wireless LAN IEEE802.11b technology.