WO2000070821A2

WO2000070821A2 - Packet classification state machine

Info

Publication number: WO2000070821A2
Application number: PCT/CA2000/000580
Authority: WO
Inventors: Feliks J. Welfeld
Original assignee: Solidum Systems Corp.
Priority date: 1999-05-18
Filing date: 2000-05-17
Publication date: 2000-11-23
Also published as: JP2003500709A; EP1190314A2; AU4739500A; WO2000070821A3; IL146539A0

Abstract

A method of reprogramming classification data in a packet classification state machine without interrupting the operation of the state machine is disclosed. Data relating to a plurality of new nodes from a starting node of the classification tree within a classification tree are stored such that they accurately indicate subsequent nodes within the existing data structure. Once the data is stored, a new first node address is stored in a predetermined location. Thereby causing subsequent state machine executions to begin at a new node. Preferably, the new first node address is stored using an atomic operation such that no reading of the first node address is possible during the store operation. The method allows a plurality of state machines to simultaneously use a same classification data memory because the method does not involve overwriting existing data.

Description

Packet Classification State Machine

Field of the Invention

The invention relates to programmable state machines and more particularly to programmable packet classification state machines for use in high-speed communication.

Background of the Invention

In the past, programmable state machines were provided with procedural programming. This permits flow from one state to another state upon certain conditions according to a desired sequence. Because a state machine effectively controls a sequence of processing steps, the concept of a truly non-procedural programming method for a state machine appears contradictory. To this end, many "non-procedural" methods of programming state machines have been proposed. Nevertheless, upon analysis these have been found to be procedural or somewhat procedural. The concept of branches and loops is common in these methods and a flow diagram of operation is often evident.

Until now because of the procedural nature of state machines, non-procedural methods of programming the state machines have been regarded as having few advantages. Some of the "non-procedural" methods of programming a state machine focus on reuse of programming code and modularity of the code. Therefore, procedural elements within each state are not only permitted, but are viewed as somewhat essential. Different states, however, are commonly easily reused or assembled into a whole classification description.

When reprogramming a state machine, it is important to consider a concept of a truly non-procedural method of implementing a programmable state machine. Changes to state machine programming are then added without concern for state flow, since the description of the state machine is non-procedural. Further, implementation of the state machine is performed by the compiler and does not require significant design by a state machine designer or, more importantly, an end user.

Reprogramming of a state machine having procedural programming is considered a straightforward task. State machine execution is paused, new programming is loaded into program memory and then the state machine is restarted. The process of pausing the state machine often involves halting data flow, which is undesirable. For use in firewalls and other security applications, a change in programming often results from a change in security procedures. As such, it is important to load the change as quickly as possible. Of course, it is evident to one skilled in the art that it is often impractical to shut down a system during reprogramming and, hence reprogramming of a system often only occurs at certain times. This is inconvenient.

Another problem is encountered with current reprogramming techniques for programmable state machines when several state machines share a common program memory. When upgrading the programming for any of the state machines, all state machines are paused. This is a significant problem that is possibly avoided by duplicating program memory in order to allow all but one state machine to remain operative during a reprogramming operation. Another method of avoiding this problem is to provide each state machine with dedicated state machine memory.

A current area of research in high-speed state machine design is the area of digital communications. Commonly, in digital communication networks, data is grouped into packets, cells, frames, buffers, and so forth. The packets, cells etc, contain data and classification information. It is important to classify packets, cells, etc. for routing and correctly responding to data communications. An approach to classifying data of this type uses a state machine.

For Gigabit Ethernet, it is essential that a state machine operate at very high speeds to process data in order to determine addressing and routing information as well as protocol-related information. Unfortunately, at those speeds, memory access is a significant bottleneck in implementing a state machine or any other type of real time data processor. This is driving researchers to search for innovative solutions to increase classification performance. An obvious solution is to implement a classification state machine completely in hardware. Non-programmable hardware state machines are known to have unsurpassed performance and are therefore well suited to these high data rates; however, the implementation of communication protocols is inherently flexible in nature. A common protocol today may be obsolete in a few months. Therefore, it is preferable that a state machine for use with Gigabit Ethernet is programmable. In the past, solutions for 10 Mbit and 100 Mbit Ethernet data networks required many memory access instructions per state in order to accommodate programmability. This effectively limits operating speeds of the prior art state machines.

A programmable state machine for classification of data can be implemented entirely in software. Of course, software state machines are often much slower than their hardware equivalents. In a software state machine, each operation is performed by a software instruction and state changes result in branch operations. As is evident to those of skill in the art, to implement a high-speed state machine in software for packet classification, requires many instructions per second - many more than a billion - requiring expensive parallel processors or technologies unknown at present. In fact, a severe limitation to performance is the speed of memory devices. For example, should a 7 ns memory device be used, less than one memory access per memory device is possible for each bit of a Gigabit Ethernet stream. Thus, if each byte - 8 bits - of data is processed in a single state, only one memory access operations is possible for each state. To implement such a system as a purely software solution is impractical and unlikely.

Current state of the art integrated memory devices achieve performance in the area of 7 ns per memory access when timing and other factors are taken into account. Therefore, pure hardware implementations of state machines fast enough to implement a Gigabit Ethernet packet classifier are possible so long as only one memory access is required for every 8 bits within the Ethernet data stream. Prior art implementations of such a state machine use a branching algorithm to allow state transitions within the time frame of a predetermined number of bits. The address data for the branching algorithm is stored in program memory. When the predetermined number of bits is 8, each state transition occurs within 8 ns. One method of achieving this performance is to store a table of data having 256 entries for each possible state. The table address is then concatenated with 8 bits from the data stream to determine a next state address. This continues until a value indicative of a classification or a failure to classify is encountered.

Unfortunately, the amount of memory required to implement a system, such as that described above, is prohibitive. For example, using 8 bits at a time requires up to 256 entries per table, 16 bits at a time requires 65,536 entries. The exact number of tables also depends on a number of terminal states. Since integrated memory having a high storage capacity is not available, implementation of a prior art programmable packet classification state machines having large numbers of edges in integrated memory is currently not feasible.

It has been found that a programmable hardware state machine for use in packet classification of high-speed data communications wherein reprogramming of the state machine is performed according to a non-procedural method, would be advantageous.

It would also be advantageous to provide a method of programming a state machine such that easy modification of state machine programming is possible during operation of a plurality of state machines using the same program memory.

Object of the Invention

In order to overcome these and other limitations of the prior art, it is an object of the invention to provide state machine architecture for supporting a plurality of state machines simultaneously.

It is another object of the present invention to provide a method of reprogramming state machine program memory without requiring disruption of state machine operation.

Statement of the Invention

In accordance with the invention there is provided a method of programming state machine memory, the state machine memory comprising a plurality of storage locations including a first state address storage location, comprising the steps of: providing data corresponding to each new state, the data including data corresponding to all states preceding each new state including a new first state; storing the data within the program memory, data relating to each new state stored at a state address for said new state, the data stored in unoccupied storage locations; storing data relating to the new first state at the new first state address, the data stored in a storage location unoccupied by current state machine programming; and once the data corresponding to each new state is stored, replacing data within the first state address location with the new first state address.

In different embodiments of the invention programming of the state machine is performed during execution of the state machine, and programming of the state machine is performed without interruption of execution of the state machine. In order to accomplish this, the first state address location can be read at any time and will return proper data results. A method of achieving this is to store data within the first state address location atomically. Another method is to provide two registers for the first state address, an active register and an inactive register, activating the inactive register being performed atomically.

In an embodiment the state machine memory is for simultaneous use by a plurality of state machines. According to this embodiment the programming of each of the plurality of state machines is optionally same programming or some state machines have different programming. Optionally, programming of a state machine from the plurality of state machines is performed during execution of the plurality of state machines without interruption of execution of any other state machine from the plurality of state machines; programming of a state machine from the plurality of state machines is performed during execution of the plurality of state machines without interruption of execution of any state machine from the plurality of state machines; and/or programming of a state machine from the plurality of state machines, data relating to another state machine from the plurality of state machines remains unaltered.

In accordance with the invention there is also provided a method of programming state machine memory, the state machine memory comprising a plurality of locations and a first state address storage location, comprising the steps of: providing an image of current state machine memory on a computer; altering the state machine programming; determining states that are modified within the current state machine; determining states preceding the states that are modified including a new first state; determining unoccupied memory locations within the current state machine memory; compiling the modified states and the states preceding the states that are modified for storage in unoccupied memory locations within the current state machine memory to form reprogramming data; storing the reprogramming data within the unoccupied memory locations within the current state machine memory; once the reprogramming data is stored, replacing data within the first state address location with the new first state address; and, updating the image of the current state machine.

In an embodiment the method also comprises the step of: determining an amount of unoccupied memory; determining if the reprogramming data will fit within the unoccupied memory; and, when the reprogramming data will not fit within the unoccupied memory, compiling a portion of the modified states and the states that precede them to form new reprogramming data and replacing the reprogramming data with the new reprogramming data.

Preferably, the program memory is only read by the state machine. This is achieved by storing a current state address in memory outside the program memory. In such an embodiment, only the reprogrammer writes data to the program memory. This is advantageous since, when a plurality of state machines operates from a same memory, a state machine is assured of state machine data integrity.

In accordance with another aspect of the invention there is provided a reprogrammable state machine memory for simultaneous use by a plurality of state machines, the reprogrammable state machine memory comprising: a program memory for storing data relating to states within each state machine, the data for each state stored at a state address; a plurality of initial state address memory locations, each initial state address memory location for storing a first state address of a state machine from the plurality of state machines, means for storing the data within the program memory, data relating to each new state stored at a state address for said new state and data relating to the new first state stored at the new first state address, the data stored in storage locations unoccupied by any of the plurality of state machines; and, means for once the data corresponding to each new state is stored, replacing data within the first state address location with the new first state address.

Preferably, the plurality of initial state address memory locations is a plurality of registers, one register for each state machine from the plurality of state machines.

Accordingly, the invention provides a packet classification state machine for classifying data from a data stream. The state machine comprises: a) a programmable memory for storing information relating to states within the state machine, the states including a first group of states and a second group of states, the first group of states each represented by a table of data at a table address and including a first plurality of table elements addressable at an offset from the table address, each of the table elements indicative of a next state within the state machine, and the second group of states each represented by a table of data including a table element and occupying less memory than a table of data representing a state of the first group of states; and, b) a processor for determining a next state based on contents of a table element of a present state at an offset from the table address, the offset determined during a table address load portion of an instruction cycle in dependence upon the bits in the data stream, the processor also for switching the state machine into the next state so determined.

According to another embodiment, the invention provides a packet classification state machine for classifying data from a data stream. The state machine comprises: a) a programmable memory for storing information relating to states within the state machine, the states including three groups of states; the first group of states each represented by a table of data including 2" table elements addressable at an offset from the table address, the table elements indicative of a next state within the state machine, n bits in the data stream for determining the offset, the second group of states each represented by a table of data including 2ⁿ table elements having a size smaller than that of table elements of the first group of states, the table elements addressable at an offset from the table address, the table elements indicative of a next state within the state machine, n bits in the data stream for determining the offset; and the third group of states each represented by data indicative of a single possible next state; and, c) a processor for storing for a table address of the first group a current state address and a plurality of bits from the data stream together to form an address, for a table address of the second group a current state address and a plurality of bits from the data stream together to form an address, and for a table address of the third group a current state address; for retrieving from that address in the programmable memory an operation, the operations comprising a jump operation including data relating to a next state address; for determining the next state based on one of information within the operation; and for switching the state machine into the next state, wherein one and only one operation retrieving information from the programmable memory is performed between successive state transitions.

According to another aspect of the invention, there is provided a method of packet classification for classifying data from a data stream. The method comprises the following steps: a) providing classification data comprising information relating to states within a state machine, the states including a first group of states and a second group of states, the first group of states each represented by a table of data at a table address including a first plurality of table elements addressable at an offset from the table address, each of the table elements indicative of a next state within the state machine, and the second group of states each represented by a table of data including a table element and occupying less memory than a table of data representing a state of the first group of states; b) providing a table address of a current state; c) selecting a table element, the table element selected in dependence upon the table address, the table contents, and an offset based on a plurality of bits in the data stream; d) determining a next state of the state machine based on the content of the selected element of a present state; and, e) switching the state machine into the next state so determined.

Brief Description of the Drawings

An exemplary embodiment of the invention will now be described in conjunction with the attached drawings, in which:

Figures la and lb are simplified state diagrams for classification state machines according to the prior art;

Figures lc is a simplified state diagram for classification state machines according to the invention;

Figure 2 is a simplified flow diagram of a method of reprogramming a state machine during execution according to the invention;

Figure 3 is a simplified flow diagram of a method of reprogramming a plurality of state machines during execution according to the invention; Figure 4 is simplified state diagram of a greatly simplified state machine;

Figure 5 is simplified state diagram of the greatly simplified state machine of Figure 4 with modifications thereto;

Figure 6 is a simplified state diagram of the combined state machines of Figures 4 and 5;

Figure 7 is a simplified flow diagram of a method of memory recovery according to the invention;

Figure 8 is a simplified state diagram of a cyclic state machine for explanation of the present inventive method;

Figs. 9a and 9b are simplified state diagrams for classification state machines according to the prior art; Fig. 10a is a simplified packet descriptor for classifying a packet as one of four classifications - A, B, C, or D; Fig. 10b is a simplified diagram of a classification tree for the packet classifications of

Fig. 10a;

Fig. 10c is a classification tree equivalent to that of Fig. 10b shown using 2 bits per symbol; Fig. lOd is a classification tree equivalent to that of Fig. 10b shown using 3 bits persymbol;

Fig. 11 is a table indicating memory usage and opcodes for an exemplary embodiment of a state machine according to the invention;

Fig. 12a is a simplified memory diagram for a state machine memory according to the invention;

Fig. 12b is an address table for a classification tree implemented for the classification tree of Fig. 10b;

Fig. 12c is an address table for a classification tree implemented for the classification tree of Fig. 10c; Fig. 12d is an address table for a classification tree implemented for the classification tree of Fig. lOd;

Fig. 13 is a simplified block diagram of an integrated circuit implementation of an acyclic classification state machine according to the invention; and

Fig. 14 is a simplified diagram of a system according to the invention of implementing a plurality of state machines using a single same programmable classification data memory.

Detailed Description of the Invention

As used herein, the term data packet encompasses the terms buffer, frame, cell, packet, and so forth as used in data communications. Essentially a data packet is a grouping of data that is classifiable according to a predetermined classification. Classifications are commonly codified by standards bodies, which supervise communication standards.

In the state diagrams presented in the Figures, a state represented by a rectangle is a terminal state. Other than terminal states are represented by triangles. Referring to Figure la, a simplified state diagram for a state machine supporting a single data access operation per data cycle is shown. Each state is in the form of a look-up table at a state address. The states are represented as blocks having four state transitions, each shown as a line to another state. Sometimes, several state transitions lead to a same next state. A predetermined number of bits (2 for the simple diagram of Fig. la) is loaded into the lowest order bits of the address, and data from the newly formed address is read. When the data comprises another state address, the next predetermined number of bits from the data stream is loaded into the lowest order bits. Otherwise an action such as filtering or packet classification is performed.

Generally, in order to reprogram such a state machine it is necessary to interrupt operation of the state machine, which results in unwanted "downtime". New programming is stored in the state machine memory after which state machine execution proceeds. Unfortunately, data flow does not stop when filtering or classification is unavailable. In light of the above, it will be evident that interruption of a single state machine's processing capabilities has negative implications. Therefore, a method of reprogramming the state machine without interrupting the operation of the state machine is desirable.

For example, when such a programmable state machine is implemented in hardware, the address of block 1 is fixed. Typically, the address is 0. Of course another fixed address is possible. Since writing a single block, for example 1. often requires storing data in each of several locations thereby requiring more than a single clock cycle, there is a time during reprogramming when the block is partially reprogrammed. If the state machine reaches the block while it is partially programmed, indeterminate results occur. This is undesirable.

Referring to Figure lb, three state machines stored within a single memory are shown. Each has completely separate blocks and, as such, each is independent of the others. In order to reprogram one of the state machines, it alone needs to be interrupted. Unfortunately, there is little benefit in storing programming of several state machines within a single memory when they are truly independent. In contrast, where a plurality of state machines share a single memory and share some programming, reprogramming of one of the plurality of state machines results in "downtime" for all the state machines that share the single memory. Referring to Figure lc, shown is a similar state diagram to that of Figure lb in which the three state machines share a same program memory and similar blocks, for example 6, 6b, and 6c, are stored as a same block (6). As is evident from Figure lc, each state machine executes a different programming. Presently, it is difficult to reprogram the state machine memory of Figure 1 c without interrupting execution of all three state machines.

Referring to Figure 2, a simplified flow diagram of a method according to the invention is shown. A state machine executes a classification function. Preferably, the classification function is an acyclic classification function, but this need not be so. The state machine is based on a table look-up for each state transition and information relating to each state transition is stored as a table of data at a state address. In operation, a first state address is read from a first state address storage location. The first state address is an address that is read at a start of state machine operation. The first state address indicates a first table.

A programmer of the state machine comprises a processor for differentiating between storage locations that contain state machine data and those that do not. Commonly, this is performed by maintaining information relating to locations where current state machine information is stored. The programmer is provided with modifications to current state machine programming. For example, a table of data relating to a second state of the state machine is modified. The programmer writes any new information to the memory in storage locations that are unused by current state machines. Unused storage locations do not contain current state machine programming data.

As is evident to those of skill in the art, in order to link the newly written states into the existing states within the state machine requires modification of some states, which is not easily performed during operation of one or more state machines. One approach to solving this problem is to stop operation of the state machine until the reprogramming is complete. Stopping operation of the state machine prevents data being processed and is therefore undesirable. According to the invention, each state preceding any modified states is also written to program memory unused by current state machine programming. The newly written states form a start of the state machine programming from a beginning of state machine operation until a point in the state machine programming from which no further modifications are being made. Once the data is written, the first state of the newly written data is a first state of the state machine. Therefore, by modifying the information stored in the first state address location, the newly written state data is used during a subsequent execution of the state machine - the modified state machine is executed. Preferably, this operation is performed atomically. It is possible to ensure that no first state address location is read during a non-atomic first state address write operation; however, this results in a small amount of downtime for the state machine and is, therefore, undesirable. The term "modified state" as used herein denotes states that are modified as well as those states that are newly created. It will be apparent from the above description that reprogramming of the state machine memory is now possible during state machine operation absent pausing state machine execution or with minimal pause when other than an atomic operation is used to store the first state address. Once the first state address is updated, a subsequent execution of the state machine uses the programming of the modified state machine. However, until the first state address is updated, the unmodified programming of the state machine is executed.

Of course, once the first address is updated, storage locations associated with states that are modified - state data was rewritten for those states - become unused storage locations given that data for those states is not in use by another different state machine. The resulting storage locations unused by current state machine programming are then "recovered" for use in further processes of reprogramming the state machine memory.

Typically, the first state address is stored in a first hardware register to enable fast access to the address. When a hardware register is used, an atomic operation to write the first state address is typically used. For example, the number of bits stored in a single clock cycle is equal to the number of bits within the register. Alternatively, a pseudo- atomic operation is used. For example, the write operation writes a number of bits per available cycle and those bits are all simultaneously clocked into the register once the entire address is available. Further alternatively, a second other register is written in several clock cycles and a flag bit is then updated to cause the newly written register to be read as the first state address instead of the first hardware register. Other methods of making the changeover from the first hardware register with an old first state address to a hardware register with the new first state address atomically will be evident to those of skill in the art based on the above disclosure.

Alternatively, the first state address is stored at a fixed address within the program memory from which it is capable of being loaded and into which it is capable of being stored in an atomic or pseudo atomic fashion. Preferably, the first state address is accessible in a single clock cycle.

Referring to Figure 3, a method of reprogramming a state machine during execution and according to the invention is shown. The method shown is for state machine program memory in concurrent use by a plurality of state machines.

Each state machine executes a classification function. Preferably, the classification function is an acyclic classification function, but this need not be so. The state machines, similar to the state machine described with reference to Figure 2, are based on a table look-up for each state transition and information relating to each state transition is stored as a table of data at a state address. In operation, a first state address is read from a first state address storage location. The first state address is an address that is read at a start of state machine operation of each state machine. When state machines are individually programmable, each state machine has an associated first state address.

A programmer of the state machine comprises a processor for differentiating between storage locations that contain state machine data and those that do not.

Commonly, this is performed by maintaining information relating to locations where information of current state machines is stored. The programmer is provided with modifications to programming of current state machines. For example, a table of data relating to a second state of the first state machine is modified. The programmer writes any new information to the memory in storage locations that are unused by current state machine programming. Unused storage locations do not contain state machine programming data relating to any current state machine. According to the invention, each state preceding any modified states is also written to program memory unused by current state machine programming. The newly written states form a start of the state machines' programming from a beginning of state machines' operation until a point in the state machines' programming from which no further modifications are being made. Once the data is written, the first states of the newly written data are first states of the state machines. Therefore, by modifying the information stored in the first state address locations the newly written state data is used during subsequent executions of the state machines, i.e., the modified state machines are executed.

As noted above, it is important that the modification of the first state address occurs in an atomic or pseudo-atomic fashion. Alternatively, state machine operation is paused during writing of the first state address. Otherwise, it is possible that the first state address is accessed during writing of the first state address and that the address loaded for starting execution of the state machine is non-sensical.

It will be apparent from the above description that reprogramming of the state machine memory is possible during execution of any number of state machines. Once one of the first state addresses is updated, any subsequent execution of the associated state machine involves the modified state machine programming. Until a first state address is updated, unmodified state machine programming is executed. In this manner, when each state machine is in execution of different programming, the present invention permits reprogramming of one state machine without interrupting operation of other state machines.

Once a new first state address is stored, it is possible to recover memory. Even though memory recovery is optionally performed immediately, reprogramming should not be performed until the currently executing state machines restart. This prevents the possibility that a state of the previous state machine that is in execution is overwritten during execution. A method of memory recovery is outlined below.

With the data associated with each state a counter indicating a number of states referencing that state is provided. When the first state address register is changed, a counter associated with data at the address pointed to by the contents of the first state address register prior to the change is decremented since there is one less reference to that state data. When the counter is zero, memory locations associated with that state are recovered and counters associated with data referenced from the data in the recovered memory locations are decremented. The memory recovery operation proceeds recursively until no counters having a zero value remain.

In an alternative embodiment, a timer is provided for providing a timing signal to pause state machine execution when the timer has expired. This is useful for programming of the programmable memory where a program must be stored by a certain time but also may be stored anytime prior. The program is entered and reprogramming is performed according to the above-described method. If reprogramming is not completed within the specified time, the state machine is paused between classification operations and the programming is completed.

According to another embodiment of the invention, when a buffer is used to buffer data prior to its provision to the state machine, the timer may be implemented to determine an amount of processing time to devote to the reprogramming task. For example, when new programming is introduced, it may be desirable to provide 10% of the processing power or the state machine devoted to reprogramming. Of course, when the state machine requires less than 90% of its processing power, the reprogramming operations may consume the available processing time. A non-limiting example of an application of the present invention is shown with reference to Figures 4, 5, and 6. Figure 4 is a state diagram of a current state machine having states 1 to 11. When it is desired to modify states 6, 10, and 11, prior art devices require pausing of state machine execution. For ease of description, the state diagram is shown with eleven states, however, in practice the number of states is only limited by the memory size of the state machine and the operation performed. Alterations are made to an image of the state machine of Figure 4 resulting in a state machine represented by the state diagram of Figure 5 having modified states 6n, 12, 13 and 14. As one skilled in the art will appreciate, it is possible to modify or add many new states, the number of which is only limited by the memory size of the state machine memory. Memory contents for the state machine illustrated in Figure 4 are shown in Figure 6 along with the modified states In, 2n and 6n. Data relating to the new and/or modified states, i.e., In, 2n and 6n are written into memory storage locations unused by currently executing state machine programming. In Figure 6 the state machine programming for both state machines, that of Figure 4 and that of Figure 5, are evident within the state diagram. From state 1 a first state diagram proceeds and from state In a second other state diagram proceeds. The states 3, 4, 5, 7, 8, and 9 are common to both state diagrams. In fact, state In is similar to state 1 but since state 2n is different, it is also different. Similarly state 2n is similar to state 2 but since state 6n is different, it is also different In order to distinguish between the state machine of Figure 4 and that of Figure 5, one must determine the start address of the state machine. Once all the new and/or modified state data are stored in memory the start address of the state machine is updated with a new first state address, that of state In. The operation is performed atomically or pseudo-atomically. The next time the state machine begins execution, it uses the newly stored data. Sufficient state machine memory is provided to allow use of either state machine programming depending on the selected start address. Thus, even if three state machines use identical programming, it is possible to modify the programming of one and not the others, given three separate first state address locations. It is important that state transitions are maintained during modification of state machine programming to prevent "downtime". When the start address is that of state In, the state machine diagram of Figure 5 is the current state machine diagram having modified states In, 2n and 6n. Memory locations associated with states that are not accessed by any state machine - in this case by the single state machine - are now "free" storage locations or storage locations unused by current state machine programming.

It is evident to those of skill in the art that all memory locations are occupied and that the term "unoccupied" as used herein refers to memory locations that are not occupied by current state machine program data.

Of course, when two or more banks of memory are used, switching between banks allows for reprogramming of the off-line bank while execution of programming in the on- line bank occurs. Effectively, in some situations, the two banks of memory are program memory and the start addresses in each bank are identified by a current bank. Therefore, the indicator for identifying the current bank forms a start address storage location. When two banks are used, the start address storage location need only accommodate one bit of data which is necessarily written in an atomic fashion. Though the start address is described above as stored in a location, it is possible to devote two locations to the first state's programming and to select between the two locations based on an indicator bit. In these cases, the indicator bit forms the start address - address 0 or address 1 - and the data may be stored in a flip-flop or in another conventional fashion.

Also, though the step of loading the current state address at the beginning of state machine execution refers to loading the address from the start address location, when a flip flop is used to indicate a current bank of memory, the address of the first state's programming is typically fixed except for selection of the appropriate bank. Therefore, the current state address is loaded with the fixed address though the memory location that the programming data is retrieved from is dependent upon the content of the data relating to the start address. In this case, the current state address register contains less than the necessary address data to determine the current state address and the start address location forms part of the current state address in so far as it determines which of the available fixed start addresses to use.

Referring to Figure 7, a simplified flow diagram of a method of recovering memory unused by any currently executing state machine is shown. When new state data is stored in the program memory, the existing state data remains unchanged. Once the first state address of each modified state machine is updated to reflect a new first state address for that state machine, the replaced state machine data is identified and noted as unused state machine memory.

In order to identify unused memory locations, a value is stored with state data for each state indicating a number of references to that state data. The state data to which the previous first state address referenced is accessed and the value therein is decremented. If the value is 1 or greater, then their still exists a reference to that state data and it is not noted as unused memory. If the value is zero, then the state machine data at that location is not referenced and it is noted as unused. All state machine data referenced from that data is then accessed and the value associated therewith is decremented. The method proceeds, recursively, until there remains only state machine data with associated values of 1 or more. Of course, the method could also be implemented iteratively if so desired. The use of such a method makes memory recovery simple and allows for self contained state machine programming including all necessary information for programming and reprogramming of same. It is therefore advantageous.

Though states are described as preceding other states or following other states, this terminology is most applicable to an acyclic state machine. For cyclic state machines, the term "reachable from" is a more accurate statement of a state that follows another state. For example, state A is reachable from state B is, for an acyclic state machine, equivalent to state A follows state B. A cyclic state machine is shown in Fig. 8 in which A is reachable from B. Incremental programming according to the invention is still useful. For modifying a single node all nodes from which the modified single node is reachable are then updated. If from a part of the graph the updated nodes are not reachable, that part of the graph need not be modified.

Referring to Fig. 9a a simplified state diagram of a typical classification state machine according to the prior art is shown. Transitions between states are represented by a series of operations shown as lines connecting states shown as circles. The state diagram is for an acyclic state machine and each state is followed by one of a number of possibilities. Such a state machine is easily implemented in either software or hardware. Unfortunately, as the speed of the state machine operation is increased, operations for each state transition must be executed concurrently in order to achieve necessary performance. This bottleneck has, heretofore, required dedicated non-programmable hardware state machine design. In Fig. 9b, a reduced state diagram is shown for the state machine of Fig. 9a. Here, terminal states of the classification A are combined into a single state 901a. Other states are similarly combined. As is evident from the diagram, the number of edges representing state transitions is somewhat reduced - from 20 to 16. Some states, such as state R, are terminal states, ACCEPT or REJECT. Restarting the state machine follows these states. The restart typically occurs before the beginning of the subsequent packet. Typically, there is a means external to the present invention to identify the start of each packet.

Referring to Fig. 10a, a simplified diagram of a greatly simplified protocol for packet classification is shown. The simplified protocol is used to facilitate understanding of the invention absent detailed knowledge of Ethernet or other communication protocols. Four bit patterns are shown, each representing a different classification. The bit patterns are similar. A first set of three bits must each be one or the data within the data stream remains unclassified. This is followed by 8 bits that are not important to the classification excepting that they occur. Three more bits must each be one and then eight more "don't care" bits. The final two bits are then used to distinguish between the four classifications. In Fig. 10b, a classification tree for implementing a programmable state machine is shown. A typical packet classification tree comprises data relating to a plurality of classification protocols each of which has many bits; the classification tree shown in Fig. 10b is simplified to facilitate explanation of the invention. Typical classification trees result in a very large data structures that are, in many instances, too large to store in a single integrated memory device.

For a gigabit Ethernet a single bit arrives in Ins. For a gigabit Ethernet packet classification state machine operating on three bits per state, the first three bits arrive in 3ns. Once provided to the state machine, there is a lag of 3 ns until a further 3 bits arrive. During those 3 ns, all operations for a state transition are completed. Of course, when more time is required to prepare for a state transition, more bits are grouped together. Thus, depending on the speed of communications, a state machine for classification using two bits at a time is possible as is one using 8 bits or 16 bits at a time. Using embedded memory within the programmable device, fast memory access operations are possible, therefore for eight-bit operation - 8ns per state transition - at most 2 sequential memory access operations from a same programmable memory, and more likely only a single memory access operation, are possible for each state transition.

Such a state is easily implemented in a look-up table having a number of addresses or in another form of address conversion. For a look-up table implementation upon receiving the 8 data bits, only a single memory access is required for determining the next memory address for the look-up table. As an example, a current table address is loaded in the high order bits of a register. 8 bits from the data stream are loaded into the low order bits of the register and act as an offset of 0-255. Once loaded, data at the location indicated by the register is loaded into the higher order bits. It is checked for an accept or reject and the next 8 bits from the data stream are loaded into the low order bits to form a new address. This continues until a terminal state is reached.

As described above for large numbers of bits, look-up table storage becomes very large and results in increased costs and reduced performance by forcing the use of external memory devices. Therefore, it is evident that there is a practical limitation to the number of bits that may be processed in parallel in a programmable state machine. Also, since memory circuitry capable of supporting necessary speeds is currently limited to integrated memory circuitry, there is a limitation on the amount of memory that is available as classification data. Because of this limitation, memory optimization provides significant advantages.

For packet classification state machines, the resulting state is of significance for determining an operation or a packet class. As such, packet classification state machines follow a tree structure. When each node of the tree relates to one bit as shown in Fig. 2b, the tree is very large and has many nodes, each node having two child nodes and one parent node. When each node relates to 8 bits, the tree has far fewer levels, but actually has a similar number of edges. In Figs. 10c and lOd two bits and three bits, respectively, relate to each node. As is evident to those of skill in the art, many nodes are of no consequence and, hence, optimization of those nodes, which are useful in packet classification is an important aspect of a programmable state machine for this purpose.

A preferred method of classification data optimization makes use of novel memory optimization techniques in order to provide reduced classification tree data having same information therein. Referring to Figs. 10b, 10c, and lOd, alternative representations of the same classification tree are shown. As is evident, there are a number of nodes that require more or less information than others. This is used, according to the present invention, in order to optimize memory storage requirements. For example, storing each node having only one possible next node - each edge from the node is to a same destination node - as a single edge reduces the number of edges in Fig. 10b from 50 to 34. This is a significant reduction in overall storage requirements. Of course, using a different number of edges per node, results in different savings. As examples, the same optimisation technique reduces the number of edges in the tree of Fig. 1 Oc from 48 to 27 and in the tree of Fig. lOd from 64 to 36.

The preferred method also groups nodes into several categories each requiring a different amount of storage - full width nodes; half width nodes; full width don't care nodes, half width don't care nodes, and so forth. By having different nodes occupying different amounts of memory, memory requirements are significantly reduced. Further, the representation of data as one of the groups is easily distinguishable. Where four groups exist, two bits are sufficient to classify an element into a group. Memory optimization is preferred since, as disclosed above, once external memory is used performance is decreased. Of course, other forms of optimizing memory usage are also applicable in association with the present invention.

For implementation of a tree traversal method according to the invention, it is preferable that every edge extending from a node - every address - is individually addressable whether it is a full width edge or a half width edge and whether it is word aligned or not. For example, full width nodes are word aligned and require, for a 2ⁿ way node, 2^π words.

Half width nodes are either word aligned or half word aligned. This allows a consistent concatenation of bits from the data stream for full width and half width nodes where, for half width nodes, a bit selects between right half of a word and left half thereof - half word aligned and word aligned, respectively. Of course, for full width nodes, no extra bit is necessary. In the diagrams of Fig. 12, half width nodes are shown as 2ⁿ half words aligned either left or right within a word. Alternatively, half width nodes are word aligned - for a two edge pair - and, for a 2ⁿ way node, require 2^n"' words.

In order to support speeds required for Gigabit Ethernet, it is preferable to handle all bit manipulation in a single stage. In order to accomplish this, each edge includes information distinguishing the group of the node to which the edge extends. This additional information includes the two bits described above and, for half width nodes, a bit indicative of right half word or left half word. A half word is referred to herein as a byte regardless of its size. Since half width nodes are byte aligned and, a word is read for each state transition, it is useful to select the appropriate byte for a current node. A preferred method of doing so is described below with reference to Fig. 12.

In an embodiment of a packet classifier with substantially optimized memory usage and capable of supporting high-speed packet classification, the following edge contents of a classification tree are used: ACCEPT, REJECT, and JUMP. Of course, other operations are possible. Some operations may require more or less data bits and therefore may necessitate other groups of nodes.

ACCEPT indicates that a packet is classified - a leaf of the classification tree is reached - and that appropriate action for the classified packet is desired. Some forms of appropriate action include passing a classification tag to a system or user, passing the packet to a predetermined routine, pushing the packet onto an application stack for a known application, and so forth. Packet processing is well known in the art of Ethernet communication.

REJECT indicates that a packet is not of a classified packet type. As such no "appropriate" action is performed, though a default action is desirable in many situations. For example, a reject tag is provided to a system or user to indicate that classification was unsuccessful.

JUMP affects the contents of a current status register. In effect, this operation results in a change of the current node and therefore of the current state. A JUMP operation loads an address contained in the word with the JUMP operation into the status register. This results in a change of state - a change to a different node within the classification tree. As such, a memory access is performed to retrieve the edge information and a further register to register transfer is performed to load new contents into the status register. Using a packet classification state machine designed specifically to implement the present invention, these actions are easily performed within existing time constraints. The inclusion of a command and data required to complete the command within a same word of data enables this performance.

Referring to Fig. 1 1 , a table shows memory word contents. Each word in a full width node comprises 2w+2 bits, where w is sufficient number of bits to represent an address within the classification tree - within the state machine. As is evident from the table, other than the ACCEPT command, all other commands fit within a half word. Alternatively, some memory addresses are represented by more than w bits and some JUMP operations are full width. This is significant for state machine optimization. As shown, the width need not include the lower order bits, which are inserted from the data stream being classified.

It is noteworthy that when half width node data is stored contiguously - two table elements in a same table forming a single word - the lowest order bit from the data stream indicates right or left half word and, therefore, until that bit is loaded, setup of the bit shifting or multiplexing necessary for loading the desired half word can not commence. Alternatively, when all edges relating to a given node are stored in either the right or left byte, the lowest order bit, other than those inserted from the data stream, of an address of a half width node is indicative of a half word left or right - high order or low order respectively. In order to facilitate operation of the state machine, it has been found that storing all edges relating to a given node in either the right or left byte and distinguishing right and left using a bit within the node data instead of the lowest order bit, allows set up of the data path to occur before data is read and therefore improves overall performance.

Preferably, an indication that a node is half width, full width, or don't care is registered and stored for use in the subsequent state. Stored in memory is node information of a direct acyclic state machine. Preferably, protocol descriptions are provided and during configuration, those protocols that are selected are compiled into a classification tree, optimised, and stored in programmable memory of the state machine device. Nodes of the state machine have two of the following formats resulting in four possible formats: full width/half width, and don't care/ 2ⁿ way. Each of these node formats is related to a format for storing edge information extending from the nodes. Don't care nodes have one edge to a following state node. 2" way nodes have 2ⁿ edges. Full width nodes are implemented such that each edge is allocated a complete word, while half width nodes store information relating to each of two edges within a single word. Basically, some edge information requires a half word of information or less and others require more. When a single edge from a node requires more than a half word of information for implementation, then all edges from the node are full width. Otherwise, the edges are optimized to half width. Also, where all edges from a node are identical, storage is reduced to a single edge. Of course, other node types having other memory storage optimizations are possible within the scope of the invention.

Referring to Fig. 12a, a memory map of packet classification data memory according to the invention is shown. The memory is divided into four areas. The areas support the four formats of nodes - don't care nodes, both full width and half width, and 2" way nodes, both full width and half width. Each JUMP instruction has two additional bits to identify the type of the next node - don't care/2ⁿ way and full/half width. For half width nodes, another bit indicates the alignment as word aligned or right aligned (half word aligned).

Optimization of the data representation of a state machine according to the invention is described below with reference to the classification tree diagrams of Figs. 10b, 10c, and lOd. Resulting memory contents are shown in Figs. 12b, 12c, and 12d, respectively. Because the classification tree data structure is so small, it is difficult to see the memory savings that result. For example, looking at Fig. 4d, it appears that significantly more memory is necessary for the 2 way tree than for a 2 way tree. When analysed, it is evident that much of the memory space is unused. A large tree would likely use the space more efficiently. Also of note, though the overall number of nodes is greatly reduced in the tree of Fig. lOd, the corresponding memory requirements are similar.

Without memory optimisation, the tree of Fig. 20b requires storage for 50 edges and the tree of Fig. lOd requires storage for 64 edges. Therefore, maintaining similar memory storage requirements for a tree having more edges per node is advantageous.

In each of the tables of Figs. 12b, 12c, and 12d, half width don't care nodes are situated above the dark line and half width 2ⁿ way nodes are below the line. Since the first node is a 2ⁿ way node it is shown at the first address on the left (word aligned) below the dark line in each of the tables. Also, when ACCEPT operations are full width, a full width node section of memory would exist. Here, since only four classes exist, half width ACCEPT instructions are supported. Referring to Fig. 12b, the state machine begins at address OOlOOx. The lowest order bit, shown as an x, is concatenated from the data stream. This results in an instruction - here a Reject or a JMP. The instruction is retrieved and executed.

A simplified block diagram of a classification state machine according to the invention is shown in Fig. 13. The state machine is useful for classifying data from a data stream and in particular for classifying data packets. The state machine, as shown, is integrated within a single integrated circuit 1 100. The integrated circuit 1 100 comprises a programmable memory 1110, a processor 1130, and a programmable memory arbiter 1140.

The programmable memory 1110 in the form of fast static random access memory

(RAM) is preferred, since static RAM is capable of achieving performance speeds that enable support of high-speed operation. Of course for low speed operation, such as that necessary for operating a 10 Mbit Ethernet packet classification state machine, known DRAM circuitry is used to increase densities and decrease costs. The RAM 1110 is for storing information relating to classification of stream data - classification tree data. Alternatively stated, the RAM 1110 is for storing data relating to states within the state machine.

For example, in the RAM 1110 data relating to each state machine node is stored. The data is stored in tables, each table having a table address and a table format. The RAM 1110 is accessible by the processor 1 130 one time per state transition. Also, the processor 1130 has a highest priority when accessing classification data within the RAM 1110. This ensures that no other RAM access operations affect state machine performance. The programmable memory arbiter 1140 controls access to the RAM 1110. The programmable memory arbiter 1 140 ensures that reprogramming of the RAM 1110 does not effect performance of the state machine. Effectively, at low speeds the arbiter 1140 guarantees the processor a single access to the RAM 1110 for every state transition. For faster data rates, the arbiter prevents RAM access operations other than by the processor 1130 while classification operations are underway. This essentially limits reprogramming of the programmable memory 1 1 10 to times when the state machine is disabled to allow for reprogramming or when incoming data is part of a classified packet, there are no packets requiring classification. A state machine is commonly disabled to allow reprogramming when it is used in security applications requiring changes in programmable memory contents due to security concerns.

The processor 1 130 is for retrieving information from the programmable memory 1110 one or fewer times per state transition. Preferably, a single access to programmable memory occurs for each state transition during classification. Of course, once data is classified, there is a pause in state machine operation until a next packet commences. During this pause, the RAM arbiter 1140 allows programming of the RAM 1 1 10. Thus, even at very high speeds, the state machine is fully programmable.

The processor retrieves data from the RAM 1 1 10 and uses the data to perform an operation for determining a next state and then switches the state machine into the next state so determined.

Though herein the ACCEPT operation is provided with a small amount of bits for indicating a classification, this is optionally increased in any of several ways. In a first way, an ACCEPT operation is implemented with a value that is an index to a classification, thereby allowing lengthy classification codes. Alternatively, ACCEPT operations are implemented using full width nodes providing an extra w bits for classification results. Of course, the two methods may be combined.

Preferably, the arbiter 1 140 provides transparent access to the programmable memory 1 110 for the processor 1130. This is achieved by blocking memory access operations for reprogramming the program memory 1 110 when a memory access from the processor 1 130 is a potential operation. Since the processor 1130 is provided highest priority, it is essential that it not be blocked in attempting to access the programmable memory 11 10. Alternatively, other methods of arbitrating between ports are employed.

Of course, since the processor does not access programmable memory 1 110 more than one time per state transition, when more than a single RAM access is possible per state transition, the processor is allocated one access to programmable memory 11 10 and reprogramming the programmable memory is performed during other accesses to the memory for a same state transition. When new packet data is detected, the processor is provided with a start address for beginning classification operations. The start address is programmable and is used in accordance with a method of reprogramming the programmable memory as described below.

Clock sources for use according to the invention are selected based on an application and based on current knowledge within the art. Selection and implementation of clock sources and of methods of synchronizing clocks to network provided clock sources are well known in the art of communication hardware design.

Unlike conventional solutions for Gigabit Ethernet that depend upon either a fast host CPU, or an imbedded CPU to examine each frame at or near real time, the present invention scales easily to gigabit rates. A CPU 10 times faster per port is not economical. Further, required processing speed varies with complexity of classification criteria. These limitations are not present according to the present invention.

In an alternative embodiment, when implementing a 10Mbit Ethernet packet classification state machine, several processors use a same programmable memory 1110. This is shown in Fig. 14. The resulting system has a plurality of processors 1 130, and a single programmable memory 1110. Each state machine is simultaneously classifying data according to a different classification tree structure. Alternatively, several processors follow a same classification tree structure. Also, classification data associated with each processor is independently programmable.

Reprogramming of the program memory 1110 is accomplished by adding new table data relating to changes in the classification tree structure of one or more state machines from an existing node in the tree back toward the root of the tree. The root address is then provided as a start address to the processor executing that state machine. When tables stored within the programmable memory 1110 are no longer used by any processor, those memory locations are reused during subsequent reprogramming. In this way, programming occurs during operation of the state machines without affecting existing programming or existing classification operations. Because of the above, it is important to automate, so much as possible, the classification data table generation and optimization process. Once automated, the specification of packet classification is unimportant in a procedural sense and becomes a process of pattern matching. Essentially, tree construction is a matter left to a packet classification program compiler. Of course, a similar system is applicable to filtering which is considered packet classification with two classes - accept and reject.

By applying memory optimization techniques in accordance with the invention, significant memory optimization occurs. This permits implementation of an integrated programmable packet classification state machine according to the invention. In order to achieve reasonable performance and cost, it is useful to optimize the table data generated by condensing tables into a smallest representation from a group of representations. Of course, where desired, other representations are used. Also, further optimization or less optimization is used depending on a particular application.

Preferably, compiler software maintains an image of the data within the programmable memory and determines memory in use and memory locations, which are no longer part of any state machine classification data. In this way, frequency of memory overflow is reduced by allowing reuse of memory once its contents are obsolete. Though this appears straightforward, because the present invention supports incremental programming and the program memory can support a plurality of processors, it is very useful to track memory allocation automatically.

Though the preferred embodiment uses a single access by the processor to programmable memory per state transition, the invention may also be implemented where unnecessary accesses to the programmable memory are replaced with, for example, a counter. Thus when four Don't care states are in succession, a count command is loaded with a count number, 4, and a next address. Data from the next address within programmable memory is retrieved once the counter has completed counting the specified number. For a single Don't care, this eliminates a single memory access.

As is evident to those of skill in the art, the programmable memory is accessed for read purposes by the processor of the state machine. The state machine processor does not write to the programmable memory. Conversely, the programming of the programmable memory only requires write operations. Therefore a memory having a write access and a separate read access - dual port - is applicable to the invention. In a further alternative embodiment, true dual port memory is used and memory arbitration is unnecessary or is provided within the memory circuitry. Alternatively, programming and diagnostics uses read write access to the programmable memory while the processor has read access to the programmable memory.

Preferably, the invention is implemented within a custom IC such as an ASIC. For example, when migrating to gigabit data rates, more efficient handling of frames rejected by the classification process is achieved by incorporating the classification system within a MAC. Once a rejection decision is made, the MAC stops DMA of the incoming frame into host memory or if DMA has not started flushes the frame from it's queue. The same receive descriptor is then used without any involvement of the driver software, thereby, reducing traffic on a host's bus.

There are several other benefits of integrating the method according to the invention inside the MAC. These include: elimination of a host bridge reducing host bus loading by loading the pattern memory through the MAC's interface; no storage required for associating classification results with a ring descriptor as required when using a separate classification component since the association is obvious; and benefits of caching are maximized because the receive ring descriptor is located anywhere that the driver desires.

Alternatively, when implemented within an FPGA, the design is preferably optimized for implementation within the type of FPGA used.

It is also possible to implement don't care bits using a don't care mask such that individual bits are "do care" bits and others are "don't care" bits. This is a straightforward issue when the don't care bits fall at either end of the table element data. In order to implement this functionality for bits in the middle of a table element data value, a shift operation is performed dependent upon the mask data in order to align the bits appropriately and adjacent one another. Because of the speed requirements of the classification state machine, such a shift operation likely requires significant hardware. Using this approach, for a 32 bit packet classification engine, a single don't care bit eliminates 2³¹ possibilities.

Numerous other embodiments of the invention are envisioned without departing from the spirit or scope of the invention.

Claims

ClaimsWhat is claimed is:

1. A reprogrammable state machine comprising: a program memory for storing programmable data relating to at least one of state machine states and state machine state transitions; and, a start address memory location supporting one of atomic and pseudo-atomic write operations of data to the entire location and for storing data indicative of a start address of state machine programming stored within the programmable memory, the data indicative of a start address reprogrammable during operation of the state machine.

2. A reprogrammable state machine as defined in claim 1 wherein the data indicative of a start address includes data indicative of an address of data relating to a first state for execution within the state machine.

3. A reprogrammable state machine as defined in claim 2 comprising a current state address register separate from the program memory and for having stored therein a current state address for a state machine during execution of the state machine.

4. A reprogrammable state machine as defined in claim 2 wherein the data relating to a first state for execution within the state machine is stored in memory that is read-only to the state machine.

5. A reprogrammable state machine as defined in claim 2 comprising: means for storing the data within the program memory, data relating to each new state stored at a state address for said new state and data relating to the new first state stored at a new first state address, the data stored in storage locations unoccupied by the state machine and during execution of the state machine; and, means for, once the data corresponding to each new state is stored, replacing the data indicative of a start address within the start address memory location with data indicative of the new start address.

6. A reprogrammable state machine as defined in claim 5 comprising: a current state address register for storing a current state address.

7. A reprogrammable state machine as defined in claim 6 wherein re-execution of the state machine begins by loading the current state address with an address based on the contents of the start address memory location.

8. A reprogrammable state machine memory as defined in claim 7 wherein the data indicative of the new start address comprises the address of the data relating to the first state for execution in the state machine.

9. A reprogrammable state machine as defined in claim 8 wherein re-execution of the state machine begins by loading the current state address with an address previously stored in the start address memory location.

10. A method as defined in claim 1 comprising the following steps: determining a state to which a first state address that is replaced points; repeating the followiong steps: decrememnting a counter associated with the determined state, and when the counter is zero, determining that data associated with that state is unoccupied; selecting a state pointed to by data within the determined state as the determined state.

11. A reprogrammable state machine memory for simultaneous use by a plurality of state machines, the reprogrammable state machine memory comprising: a program memory for storing data relating to states within each state machine, the data for each state stored at a state address; a plurality of initial state address memory locations, each initial state address memory location for storing a first state address of a state machine from the plurality of state machines, means for storing the data within the program memory, data relating to each new state stored at a state address for said new state and data relating to the new first state stored at the new first state address, the data stored in storage locations unoccupied by data forming part of any of the plurality of state machines; and, means for once the data corresponding to each new state is stored, storing the new first state address within the first state address location in an atomic or pseudo-atomic fashion.

12. A reprogrammable state machine as defined in claim 1 1 wherein an atomic or pseudo- atomic fashion is defined as in a fashion precluding an operation for reading the first state address location during the storing of the new first state address.

13. A reprogrammable state machine as defined in claim 1 1 comprising: a plurality of current state address registers, one associated with each state machine from the plurality of state machines and for storing a current state address of the associated state machine.

14. A reprogrammable state machine as defined in claim 13 wherein re-execution of a state machine from the plurality of state machines begins by loading the current state address register associated with said state machine with an address based on the contents of the initial state address memory location associated with said state machine.

15. A reprogrammable state machine as defined in claim 14 wherein re-execution of a state machine from the plurality of state machines begins by loading the current state address register associated with said state machine with an address previously stored in the initial state address memory location.

16. A method of reprogramming state machine memory, the state machine memory comprising a plurality of storage locations and a first state address storage location, comprising the steps of: providing data corresponding to each new state, the data including data corresponding to all states from which each new state is reachable including a new first state; storing the data within the program memory, data relating to each new state stored at a state address for said new state, the data stored in storage locations unoccupied by data for use with currently executing state machines; storing data relating to the new first state at the new first state address, the data stored in an storage location unoccupied by data for use with currently executing state machines; and once the data corresponding to each new state is stored, replacing data within the first state address location with data indicative of the new first state address atomically or pseudo-atomically.

17. A reprogrammable state machine as defined in claim 16 wherein atomically and pseudo-atomically is defined as in a fashion precluding an operation for reading the first state address location during the storing of the new first state address.

18. A method as defined in claim 16 wherein the state machine is acyclic and wherein a new state is reachable from a state if said state precedes the new state.

19. A method as defined in claim 16 wherein the reprogramming of the state machine is performed during execution of the state machine.

20. A method as defined in claim 19 wherein the reprogramming of the state machine is performed without interruption of execution of the state machine.

21. A method as defined in claim 19 wherein the state machine memory is for simultaneous use by a plurality of state machines, and wherein programming of a first state machine from the plurality of state machines is same as programming of a second state machine of the plurality of state machines.

22. A method as defined in claim 16 wherein the state machine memory is for simultaneous use by a plurality of state machines, and wherein programming of a first state machine from the plurality of state machines is different from programming of a second state machine of the plurality of state machines.

23. A method as defined in claim 16 wherein the state machine memory is for simultaneous use by a plurality of state machines, and wherein programming of a state machine from the plurality of state machines is performed during execution of the plurality of state machines without interruption of execution of any other state machine from the plurality of state machines.

24. A method as defined in claim 16 wherein the state machine memory is for simultaneous use by a plurality of state machines, and wherein programming of a state machine from the plurality of state machines is performed during execution of the plurality of state machines without interruption of execution of any state machine from the plurality of state machines.

25. A method as defined in claim 16 wherein the state machine memory is for simultaneous use by a plurality of state machines, and wherein during programming of a state machine from the plurality of state machines, data relating to another state machine from the plurality of state machines remains unaltered.

26. A method of programming state machine memory, the state machine memory comprising a plurality of locations and a first state address storage location, comprising the steps of: providing an image of current state machine memory on a computer; altering the state machine programming; determining states that are modified within the current state machine; determining states from which the states that are modified are reachable including a new first state; determining memory locations within the current state machine memory unoccupied by data associated with a state machine in execution; compiling the modified states and the states preceding the states that are modified for storage in memory locations within the current state machine memory unoccupied by data associated with a state machine in execution to form reprogramming data; storing the reprogramming data within the memory locations within the current state machine memory unoccupied by data associated with a state machine in execution; once the reprogramming data is stored, replacing data within the first state address location with the new first state address; and, updating the image of the current state machine.

27. A method as defined in claim 26 wherein the step of replacing data within the first state address location with the new first state address is performed atomically or pseudo- atomically.

28. A reprogrammable state machine as defined in claim 27 wherein atomically and pseudo-atomically are defined as in a fashion precluding an operation for reading the first state address location during the storing of the new first state address.

29. A method as defined in claim 26 wherein the step of determining memory locations within the current state machine memory unoccupied by data associated with a state machine in execution comprises the following steps: determining a state to which a first state address that is replaced points; repeating the followiong steps: decrememnting a counter associated with the determined state, and when the counter is zero, determining that data associated with that state is unoccupied; selecting a state pointed to by data within the determined state as the determined state.

30. A packet classification state machine for classifying data from a data stream, the state machine comprising: a) a programmable memory for storing information relating to states within the state machine, the states including a first group of states and a second group of states, the first group of states each represented by a table of data at a table address and including a first plurality of table elements addressable at an offset from the table address, each of the table elements indicative of a next state within the state machine, and the second group of states each represented by a table of data including a table element and occupying less memory than a table of data representing a state of the first group of states; and, b) a processor for determining a next state based on contents of a table element of a present state at an offset from the table address, the offset determined during a table address load portion of an instruction cycle in dependence upon the bits in the data stream, the processor also for switching the state machine into the next state so determined.

31. A packet classification state machine for classifying data as defined in claim 30 wherein table elements comprise data relating to the format of table elements of a next state.

32. A packet classification state machine as defined in claim 30 wherein each table of data representing a state of the first group of states occupies a same amount of memory and has a same number, n, of table elements.

33. A packet classification state machine as defined in claim 32 wherein each table of data representing a state of the first group of states comprises 2^N table elements wherein N is a number of bits used by the processor to determine the offset from the table address.

34. A packet classification state machine as defined in claim 32 wherein each table of data representing a state of the second group of states occupies a same amount of memory and comprises only one table element.

35. A packet classification state machine as defined in claim 30 wherein the processor comprises means for determining a size of table elements within a table and for loading an element of the determined size.

36. A packet classification state machine as defined in claim 35 wherein the means for determining a size of table elements within a table includes means for determining an alignment of table elements within data words and for loading contents of a table element of the determined size having the determined alignment.

37. A packet classification state machine as defined in claim 30 wherein each table of data representing a state from the second group of states comprises data indicative of only one next state.

38. A method of packet classification for classifying data from a data stream, the method comprising the steps of: a) providing classification data comprising information relating to states within a state machine, the states including a first group of states and a second group of states, the first group of states each represented by a table of data at a table address including a first plurality of table elements addressable at an offset from the table address, each of the table elements indicative of a next state within the state machine, and the second group of states each represented by a table of data including a table element and occupying less memory than a table of data representing a state of the first group of states; b) providing a table address of a current state; c) selecting a table element, the table element selected in dependence upon the table address, the table contents, and an offset based on a plurality of bits in the data stream; d) determining a next state of the state machine based on the content of the selected table element of a present state; and, e) switching the state machine into the next state so determined.

39. A method of packet classification for classifying data from a data stream as defined in claim 38 wherein the steps (c), (d) and (e) are performed iteratively and wherein the step of (e) switching the state machine into the next state comprises the step of providing the table address of the table representing the next state as the table address of the current state for a next iteration.

40. A method of packet classification for classifying data from a data stream as defined in claim 38 wherein the step of determining a next state of the state machine is performed, for a state in the first group of states, by concatenating the table address and the bits in the data stream to form an element address, data within memory at the element address indicative of the next state.

41. A method of packet classification for classifying data from a data stream as defined in claim 38 wherein the bits are consecutive bits within the data stream

42. A method of packet classification for classifying data from a data stream as defined in claim 41 wherein the step of determining a next state of the state machine is performed, for a state in the first group of states, by concatenating the table address and the consecutive bits in the data stream to form an element address, data within memory at the element address indicative of the next state.

43. A method of packet classification as defined in claim 38 wherein the steps (c) and (d) are performed per state transition.

44. A method of packet classification as defined in claim 38 wherein the classification data is an acyclic classification tree.

45. A method of packet classification as defined in claim 44 wherein the classification data is optimized by determining states having only one possible next state and storing them as tables representing states of the second group of states.

46. A method of packet classification as defined in claim 44 wherein some information relating to a state is a don't care operation and wherein some information relating to the previous state indicates that the next state comprises a don't care operation, wherein a the state comprises fewer table elements than a state absent a don't care operation.

47. A method of packet classification as defined in claim 44 wherein some information relating to a state is a counter indicating a subsequent number of don't care operations.

48. A method of packet classification as defined in claim 38 wherein the first group of states is represented by a table having elements with a width of m bits and the second group of states is represented by a table having elements with a width of approximately m/2 bits.

49. A method of packet classification as defined in claim 48 wherein each of the first group of states and the second group of states is represented by a table having 2ⁿ table elements wherein n is a number of bits within the data stream in dependence upon which the table element is selected.

50. A method of packet classification as defined in claim 38 wherein the information relating to states within a state machine is stored in a memory and wherein the information includes information relating to alignment of table elements within the memory.

51. A method of packet classification as defined in claim 50 wherein a table of table elements is stored within the memory with a same alignment.

52. A method of packet classification as defined in claim 38 wherein the information relating to states within a state machine includes information relating to width of table elements relating to a next state.

53. A packet classification state machine for classifying data from a data stream, the state machine comprising: a) a programmable memory for storing information relating to states within the state machine, the states including three groups of states; the first group of states each represented by a table of data including 2ⁿ table elements addressable at an offset from the table address, the table elements indicative of a next state within the state machine, n bits in the data stream for determining the offset, the second group of states each represented by a table of data including 2" table elements having a size smaller than that of table elements of the first group of states, the table elements addressable at an offset from the table address, the table elements indicative of a next state within the state machine, n bits in the data stream for determining the offset; and the third group of states each represented by data indicative of a single possible next state; and, c) a processor for storing for a table address of the first group a current state address and a plurality of bits from the data stream together to form an address, for a table address of the second group a current state address and a plurality of bits from the data stream together to form an address, and for a table address of the third group a current state address; for retrieving from that address in the programmable memory an operation, the operations comprising a jump operation including data relating to a next state address; for determining the next state based on one of information within the operation; and for switching the state machine into the next state, wherein one and only one operation retrieving information from the programmable memory is performed between successive state transitions.

54. A packet classification state machine for classifying data from a data stream as defined in claim 53 comprising: means for determining a group of a table relating to a next state.

55. A packet classification state machine for classifying data from a data stream as defined in claim 53 comprising: means for determining alignment of a table relating to a next state.