CA2126503A1 - Message header classifier - Google Patents

Abstract

ABSTRACT OF THE DISCLOSURE
The classifier device disclosed herein analyzes message headers of the type which comprise a sequence of bit groups presented successively. The device employs a read/write memory for storing at a multiplicity of addresses, an alterable parse graph of instructions. The parse graph instructions include node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for previously characterized header types. A logical processor responds to a node instruction read from memory either by initiating another memory read at a next address which, given the current state of the processor, is determinable from the current node instruction and the current header bit group or by outputting data indicating recognition failure if no next address is determinable. The logical processor responds to a terminator instruction by outputting respective header identifying data. Accordingly, for a previously characterized header type, a corresponding pattern of node instruction and a terminator instruction can be written into memory thereby enabling the device to identify the respective header type in a time which is essentially proportional to the length of the header and thus also to the time of presentation of the header. The parse graph can be updated dynamically during the operation of the classifier.

Description

- 2~2~ c) MESSAGE HEADER CLASSTFIER

BACKGROUND OF_THE INVENTION

The present invention relates generally to packet switched message handling systems and, more particularly, to a device for quickly classifying previously characterized message headers.

There is increasing interest in providing communications between disparate computer systems and even between networks of differing characteristics. Further, with the availability of very high bandwidth trunk lines, e.g., using fiber optic cables, there is increasing interest in combining traffic from a great variety of sources ~or transmission through a single trunk line.
For wide area networks, packet switching technology is widely used where information to be transmitted is broken into packets o~ data which are proceeded by headers containing information useful in routing. The header may also identify the source and the destination. Whether truly packet switched or not, most digital communication systems employ message formats in which there is an identifying header of some sort.

For large and complex networks, one standard which is being adopted is that of an asynchronous transfer mode (ATM) switch which operates to transfer data cells of standard characteristics between links in a network. Within an ATM network, routing is typically provided by setting up a path between a source and destination and then including, in the headers for the '.',' , ' : . , ' ' '.' ' ' ! ' ' ' , , ' 2 ~ ?~3 standardi~ed cells or packets, data which identifies that path to the various switches along the route. However, at various points, a relatively standardized ATM network must interface with disparate syst~ms and convert the message formats of the disparate systems into the standard ATM cell format. A typical situation where this type of interfacing i5 necessary is at a so-called bridge where a ATM network interfaces with a local area network, e.g., an ethernet or a token ring network. The job of the bridge is essentially to examine all o~ the traffic on the local area network and, if the destlnation of a given packet is across the ATMi network, i.e., outside the respective local area network, the bridge transfers that message into the ATMi network.
This process is often referred to as forwarding.

Heretofore, the analysis of message headers to determine appropriate action has been provided by writing a conventional computer program to run on a high speed, but essentially conventional, general purpose processor. The program operates to dissect and test the components of the message header and thereby determine its content. If such a generalized analysis process is performed on every arriving packet or message the time required for such processing rises rapidly as the number of possible header types increa~es and for high bandwidth systems it may be essentially impossible to conduct such processing. Further, the time required to analyze a message header using this approach, can be highly variable and unpredictable. It is therefor questionable whether this approach can be sufficiently refined to meet the demands of connecting high speed AT~i networks to highly varied traffic sources.

Among the several obj ects of the present invention may be noted the provision of a novel device for classifying message.

2~ 2 ~ ~'3 headers of tha type which comprises a sequence of bit groups presented successively: the provision of such a device which will direct selected messages to a selected one of a plurality of output paths in accordance with the contents of the message h~ader; the provision of such a device which will classify a message header in a time essentially proportional to the length of that header; the provision of such a device which operates at high speed; the provision of such a device in which the time required to classi~y a header is predictable; the provision of suc~ a device whose operation can be easily altered to accommodate new types of message headers; the provision of such a device which is highly reliable and which is of relatively simple and inexpensive construction. Other objects and ~eatures will be in part apparent and in part pointed out hereinafter.

SUMMARY OF THE INVEN?ION

The classifier apparatus of the present invention operates to classify messages in accordance with the contents of a message header which comprises a sequence of bit groups presented successively. The apparatus employs a memory for storing, at a ~ultiplicity of addresses, node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for respective header types. The apparatu~ also employs logic means responsive to a node instruction read from memory for initiating a m~mory read at a next address which, given the current state of the logic means, is determinable from only the current node instruction and the current header bit group or for outputting data indicating recognition failure i~ no next address is determinable. The logic means is responsive to a terminator instruction for outputting the respective header - ` ~
2 1 2 ~

identifying data. Accordingly, the apparatus can be made to respond to a particular header type by writing a corresponding pattern of node instructions and a terminator instruction into the memory. Preferably, the logic means incorporates registers for storing its state during jumps to subroutines.

BRIEF DESCRIPTION OF THE DRA~INGS

Figure 1 is a block diagram of an ATM switch system employing a classifier in accordance with the present invention:

Figure 2 is a block diagram of an edge card employed in the switch system of Figure l;

Figure 3 is a block diagram of a from line unit employed in the edge card of Figure 2 Figure 4 is a diagram illustrating the interfacing of a classifier of the present invention with other components of the from line unit of Figure 3; and Figure 5 is a data path dia~ram illustrating a classifier of the present invention as employed in the from line unit of Figure 3.

Correspon~ing reference characters indicate corresponding parts throughout the several views of the drawings.
, .
DESCRIPTION OF THE_PREFERRED BODIMENT

While the classifier apparatus of the present invention is not particularly concerned with the details of the network '~ 2~fi~

switching process or the conversion of disparate message formats into ATM cells, the following description is useful in setting the context within which the classifier apparatus operates and has utility. Referring now to Figure 1, an ATM switch core is designated generally by reference character 10. Switch core 10 may be considered a generalized cross-bar type switch which uses packet switching techniques to transmit ATM cells from one to another of a set of line cards 11-15, the destination line cards ~eing selected in accordance with data contained in the cell headers. The line cards 11-15 are of two general types. The line cards 11-13 are designated edge cards in the sense that they constitute the edge of the ATM network and provide for interfacing with user systems, e.gO, l`ocal area networks or mainframe computers. The line cards 14 and 15 are designated as trunk cards and serve to interface between the switch 10 and the ATM trunks.

Overall operations of the ATM switch system are typically coordinated by a host computer as designated bv reference character 17. Further, distributed control of component sub-systems is implemented by microcomputers incorporated into the various sub-systems of the overall system. These local processors are referred to hereinafter as control processors and may, for example, be implemented by a thirty two bit microcomputers, e.g., of the Motorola 68040 type. Co~munications between the host computer 17 and these various control processors can be implemented by communications through the switch itself.
In addition, certain hardware functions are implemented by means o~ dedicated micro-controllers incorporated into the subsystems.
They are often referred to hereinafter as standard microcontrollers (SMC).

2~ 26~Q~

Figure 2 is a block diagram illustrating the general ~unctionality provided in one of the edge cards 11-13. Referring to Figure 2, circuitry specific to interfacing to the particular type of external line is indicated generally by block 21. As will be understood, this circuitry will vary considerably from one type of line card to another depending upon the external system to be inter~aced. The circuitry ~or handling the message traffic between the external line unit 21 and the ATM switch is conveniently divided into four sections: a From Line Unit (FLU) 22; a To Line Unit (TLU) 24; a To Switch Unit (TSU) 26; and a Fro~ Switch Unit (FSU) 28. The edge card includes a control processor 31 of the type described previously. Control processor 31 can read and load various registers and memory locations in the various hardware sections through a data bus 32.

As indicated previously, one of the ~unctions provided is the conversion of external message formats to ATM cell formats and this functioniality is provided in the edge cards and in particular in the From Line Unit 22. The From Line Unit 22 incorporates the recognizer or classifier of the present invention together with segmentation circuitry for converting line messages to ATM cells. I~ the classifier recognizes or can appropriately classi~y the header of an incoming packet, it provides to the segmentation circuitry a set of data which characterizes the message and enables the segmentation circuitry to convert the incoming message or packet into an appropriate number of ATM cells. The ATM cells are then coupled to the switch 10 through the To Switch Unit 26 which provides appropriate interfacing and buffering. Correspondingly, ATM
cells arriving ~rom the switch 10 are interfaced and bu~fered by the From Switch Unit 28. The ATM cells themselves carry the in~ormation necessary ~or reassembly into a message format 21 2 ~r,~3 appropriate for the particular type of local line being interfaced, this reassembly being performed in the To Line Unit 24. As indicated previously, the host computer 17 c~mmunicates with the various control processors through the switch itself.
To enable this communicationl the To Switch unit 24 and the From Switch Unit 28 couple such communications to the edge card control processor 31 through bus 32.

The From Line Unit is illustrated in greater detail in Fig.
3. Packets incoming from the local line interface 21 are transferred from the bus 32 to a FIFO register 34 under the control of one of the standard micro controllers, designated by reference character 36 and referred to hereinafter as the FLP-SMC. The contents of the FIFO 34 are provided to both a classi~ier 25 constructed in accordance with the present invention and to header processor and segmentation circuitry designated generally by reference character 27 and sometimes referred to hereina~ter as a HMU (header manipulation unit). The operations of the FIFO 34, the classifier 25 and the header processor 27 are coordinated by another one of the standard microcontrollers, designated by reference character 38 and referred to herQinafter as the ALF-SMC. As indicated previously, i~ the classifier 25 can recognize a message header, characterizing data is provided to the ALF-SMC 38 which generates an appropriate ATM cell header. The ATM cell headers are then combined, in the circuitry 27, with appropriate sections of message data taken directly from the FIFO register 34. Output staging and buffering are provided as indicated at block 39 prior to coupling the formed ATM cells into ~he To Switch Unit 26. As also indicated previously, the classifier 25 utilized a parse graph memory in which instruction patterns identifying known message headers are written. This parse graph memory is 2 ~ ''9'~

designated by reference character 57 and is accessed by the classifier 25 through a memory controller 56.

By means of a slave interface unit 40 which inter~aces between the bus 32 and the memory controller 56, the control processor 31 can write to memory 57~ typically relaying parse graph patterns generated by the host processor 17.

The host processor sets up in memory 57 what may be conveniently described as a parse graph which allows the classi~ier to analyze headers in a manner which is not only fast in a general sense but also entails only a predictable time cost for any given header structure, that time beiny essentially proportional to the length of the header and thus proportional also to the time required for presentation of the header. The patterns stored in memory 57 may also be characterized as a directed acyclic graph.

Message headers are typically organized as a sequence of bit groups presented successively with the groupings being based on aither four bit quantities, i.e., nybbles, or eight bit quantities, i.e., bytes. In the particular embodiment b~ing described, message headers are analyzed four bits at a time. In general, in~ormation identifying a particular type o~ header is introduced by writing into the memory 57 a sequence of instructions which test the four bit data groups from the header successively though not necessarily in order and, based on the test results, cause a branch to a next instruction. In one sense, the recognition or classification operation may be thought of as decoding a path through the parse graph of instructions.
It should be understood, however, that the paths for dif~erent headers may utilize some common segments, e.g., arising through ":

~J~ ~ 6'i~

similar bit patterns in the headers. As described in greater detail hereinafter, commonly used testing sequences can be advantageously constructed as a form of subroutine.

It should also be understood that it is not neces~iary that there be a one for one correspondence of instructions with the nybbles in the header since the instruction set allows for skips if the analysis already performed has provided a sufficient unique identification so tha~ some or all further processing is unnecessary. For example, in the application of the ATM switch as a bridge, whole classes of messages may simply be of no interest to the ATM network. Thus, they can be effectively discarded as soon as it is known that the message falls within a class to be discarded. For this reason, the apparatus of the present invention is broadly defined as a classifier though it is often referred to as a recognizer. As also suggested earlier, it is not necessary that the bit groups be examined in order. It may be appropriate to de~er examining some of the bit groups unitl after a much later bit group in the header sequences has been examined. This out-of-order processing can be implemented using a skip instruction as defined in greater detail hereinafter.

While a read only memory could be used ~or the parse graph i~ all header patterns to be classified or recognized were preestablished, it is presently preferred that the memory 57 be implemented as a read/write memory so that the instruction sequences for classifying or recognizing the various header patterns can be updated as new headers are encountered and so that the instruction sequencés ~or currently unused types of message headers can be stripped out o~ the memory. In decoding the parse graph, the classifier 25 does not write to this memory r~

-2~26~)t~`

but only provides read addresses thereto. Words thusly read out from the memory are then provided back to the classifier which utilizes those words as instructions as described in greater detail hereinafter.

As described previously, messages obtained from the local line are buffered in FIFO 34. Four bits at a time (a nybble) are read into the classifier or recognizer engine 25. The recognizer engine 25 is a semi-autonomous machine that scans FIFO 20 looking for packet headers to parse. Once a header with the minimum numher of bytes or nybble is found, the recognizer engine initializes itself and then parses until termination.

While the detailed implementation of a classifier or recognizer engine in accordance with the present invention will depend to a great extent upon the details of the overall system within which ik will operate, it is useful to explain a particular implementation by way of example. Fig. 4 illustrates in some greater detail the interfacing of the classifier or recognizer engine 25 with the other components shown in Fig. 3.

Fig. 4 shows various o~ the signals and register values which are used in the recogni2er engine and which interact with the FIFO 20; the control microcomputer; and the parse graph memory 57. Certain of these signals, e.g., those having to do with time stamping, are not directly involved with the classifier operation and are not later referenced. In the particular embodiment being described, these signals and values are defined as follows:

nybbl e_ptr holds the packet FIFO address of the next nybble to by loaded into nybbl e . It points to the current nybble 2~2~

of the packet header on the part that the recognizer is parsing.

nybble is the value of th~ packet nybble currently addressed by nybbl e_ptr .

pgm_adr holds the next Parse Graph Memory (PGM) address to be accessed during recognition.

instrl is the primary instruction register and holds the 32-bit content of the PGM node addressed by pgm adr.

instr2 is the secondary instruction register and holds the 32-bit content of the second word of the optimized compare instruction (COMPARE_EQ_OP_2).

srr is loaded during a subroutine call (JSR) instruction and holds the base return address for subroutines. During a return-~rom-subroutine, srr, offset by a specified amount, is loaded into nxt_addr.

mcm is used by the Recognizer to indicate that the Modi~ied Compare Mode is enabled. This is used to modify the Recognizer's reaction to a compare failure during parsing.
mcm is enabled whenever a JS~ instruction is executed.

scm is used by the ~ecognizer to indicate that the Source Compare Mode is enabled. The scm bit is enabled when the instrl ~SCE.JSR~ ~ield is asser~ed during a JSR.

mcm ptr is used during Modified Compare Mode to hold the valu~ of nybble Ptr when a JSR is executed. The Recogniz~r 2 ~2 ~ 3 ~

can then use this value to reload nybbl ~ ptr in the event o~
a compare failure during the subroutine.

scf ptr is used during Source Compare Mode to capture the value of mcm ptr when the Recognizer returns from a subroutine due to a comparison failure. This capture can only occur when Source Compare Mode is enabled.

scf indicates to the 5MC that an Source Compare failure has occurred. If scf is asserted, scf ptr holds the pointer to the beginning of the field that cause the compare failure.

ts dex is the timestamp index used in aging to update the aging table. It's value is read by the SMC.

ts valid indicates that the timestamp in tx_dex is valid and should be read by the SMC.

pattern_id holds the pattern ID of the packet after analysis by the Recognizer.

done indicates to the SMC that the Recognizer has terminated parsing.

gc~ indicates to the SMC that a general comparison ailure has occurred.

Packet FIFO is a nybble-addressable FIFO loaded with packet data from the line.

Parse Graph Memory holds all parse instructions. Its data port is 32 bits wide.

2126;1~? ~

While the internal construction of the recognizer engine is described in greater detail hereinafter, it is initially useful to note that its overall operation is to perform a parsing algorithm. In general it may be noted that the parsing algorithm utilizes each instruction obtained from the parse graph memory, typically in conjunction with a nybble taken ~rom the message header, in order to generate an address for the next instruction.
The parsing algorithm can be described as follows:

/* initialize recognizer */
mcm = FALSE; // modified compare mode scm = FALSE; // source compare mode tsm = FALSE; // timestamp mode scf = FALSE; // source compare failure flag gcf = FALSE, // general compare failure flag t~ valid - FALSE; // timestamp invalid nybble_ptr = value related to current port; // point to ~irst parse nybble nybble = fifo[nybble Ptr]; // get the ~irst parse nybble pgm_adr = value related to current port; // start the parse location mcm Ptr - unknown; // don't make any assumptions scf ptr = unknown; // don't make any assumptions srr = unknown: // don't make any assumptions t~_dex = unknown; // don't make any assumptions /* process instructions */
forever begin // continue until terminate instr = content of PGM at adr; // get next instruction i~ (instr is a JSR) begin mcm z TRUE // enable modified compare mode 2~2~

mcm ptr = nybble ptr; // save in case o~ compare failure i~ (instrl[SCE] = = TRUE) // source compare mode enabled scm - TRUE;
if (instrl[TSE] = = TRUE) // timestamp mode enabled tsm = TRUE;
srr = instrl[NEXT];
pgm_adr = instrl[NEXT];
end else if (instr is an RTS) begin nxt adr = srr[21:0] ¦ instrl[OFFSET]; // explicit return if (tsm == TRUE) // is timestamping enabled?
if (instrl~TS] == TRUE) begin// timestamping is - requested ts dex = instrl[INDEX];
ts valid = TRUE: // alert ALF SMC
end end else if (instrl is compare_eq, compare_eq_op 1, compare_eq_op 2 or range format) if (comparison failure occurs) process compare failure;
else begin if (instrl is compare_eq_op_l) pgm adr = (pgm adr[21:4], 4'bOOO1);
else pgm adr - instrl[NEXT~;
end else begin process other types of instructions pgm_adr = instrl[NEXT]:

-15- 2 ~ 2 ~

end end This description is stated in a form of pseudo code which is similar to either the computer language C or the hardware description language known as Verilog.

From the foregoing it will be understood that the recognizer is in some sense a logic engine or processor. While a more detailed hardware description of the recognizer engine is given hereinafter, it should also be understood that such a processor can also be defined by its instruction set.

As indicated previously, the read/write memory 57 stores instruction words. These words are 32 bits wide and the arrangement of the fields of each instruction within those 32 bits is illustrated in a corresponding table in the following description. Each of the instructions occupies a single 32 bit word except for the Optimized Equality Check which uses two words ~rom memory. In the description of each of the instructions there is both a verhal description of the operation performed as a result o~ executing that instruction and also a so-called pseudocode description of the microcoded operations performed by the logic means 53. As is understood by those skilled in the art, pseudocode is an efficient means for communicating the essential operation of a hardware device without becoming embroiled in cletailed gate level analysis. The pseudocode examples given herein are essentially patterned after the standard Verilog hardware description lanyuage. The particular embodiment described herein implements the following instruction set~

2 ~,_ 2 Compare ~or Equality (COMPARE_EQ) _ , - ~
FIELD POS LEN DESCRIPTION
_. . _ . _,_ OPCODE 31:30 2 constant: 2'bOl . _ _ VALUE 29:264 compare value ~ . ~
MASK 25:22 4 mask value for logical and operation NEXT 21000 22 next node address for successful comparison I _ _ _ . _ -DESCRIPTION:
This is the generalized compare instruction. The nybble is logically ANDed with MAS~ and the result is ~ompared to YALUE.
If the compare su~ceeds, control is transferred to NEXT, otherwise, process compare_failure is called for further processing. Successful comparison causes nybble Ptr to be incremented by 1.

The COMPARE EQ instruction can also be used to ef~ect an explicit failure ~FAIL) and thus cause assignment of a divert id as the pattern ID. This is done by letting: MASX = 4 'bOOOO and VALUE be any non-zero value. A compare failure will always result.

CONP~A~ION:
instrl = COMPARE EQ:
nybble = fi~o~nybble ptr];
if ((MASK & nybble) == instrl[VALUE]) begin pgm_adr = instrl[NEXT];
nybble ~tr = nybble Ptr + l;
end else process compare_failure;

-S2~2~

Optimized Compare (COMPARE EQ_OP_l, COMPARE_EQ_OP 2) .. _ _ _ . =~
¦ FIELD POS LEN DESCRIPTION
i _ __ _ _ _ , OPCODE 31:29 3 constant~ 3'blll I ~ _ _ _ SPARE 28:27 2 constant:2'bOO
__ _ _ _ _ _ NUM 26:24 3 number of nybbles to check - 1 (legal range:<1 NUM<=7) _ _ . eD ~ _ .
NO 23:20 4 comparison value , N1 19:16 4 comparison value _ _ _ _ N2 15:12 4 comparison value _ N3 11:08 4 comparison value _ _ -- - --11 N4 07:04 4 comparison value N5 03:00 4 comparison value I
I _ -- ~

. _ ; . - _ . , ~' FIELD POS LEN DESCRIPTION
_ _ OPCODE 31:30 2 constant: 2'bOO
_ . _ __ .
N6 29:26 4 comparison value _ _ N7 25:22 4 comparison value __ _ __ _ NEXT 21:00 22 next node address for successful comparison l , ~ _ . _ _ _ -The COMPARE EQ OP 1 instruction word must appear on an even address. The COMPARE_EQ_OP_2 instruction must be in the word address immediately following the compare_eq_op_l instruction.
The COMPARE EQ OP_1 and COMPARE_EQ_OP_2 instructions must always appear together, even if the number of comparisons is less than ,, - , . . ~ . ,,. . ~, . . . ~ .:

r~

2~ 2~

6. In this case, COMPARE_EQ OP 2 will be used only to obtain the next node address in the case of successful comparison.

D~8CRIPTIo~:
This is a space optimized compare-for-equality instruction. A
specified number (NUM+l) of nybbles are compared with as many as eight values (nO-n7). if the compare succeeds, control is transferred to NEXT, otherwise, process compare failure is called for further processing. No masking is allowed before comparison.
Successful comparison over all nybbles advances nybble_ptr by NUM
+ 1.
instrl = COMPARE EQ OP l;
instr2 = COMPARE EQ_OP 2;
nybble = fifo[nybble ptr];
for (count=0; count<=num; count-count+l) begin case (count) begin O :if (nybble == instrl[NO]) pgm adr - instr2[NEXT];
else process_compare failure;
1 :if (nybble == instrl [Nl~) pgm adr = instr2[NEXT];
else process compare failure;
2 :if (nybble == instrl [N2~) pgm adr = instr2[NEXT];
else 21 2~

process compare_failure;
3 :if (nybble == instrl [N3]) pgm adr = instr2[NEXT];
else process_compare_failure;

4 :if (nybble == instrl [N4]) pgm_adr = instr2[MEXT];
else process_compare_failure;
S :if (nybble == instrl [N5]) pgm_adr = instr2[NEXT];
else process_compare_Pailure;
6 :if (nybble == instr2 [N6]) pgm adr = instr2[NEXT];
else process_compare_failure;
7 :iP (nybble == instr2 [N7]) pgm adr = instr2[NEXT];
else process_compare_failure;
endcase nybble ptr = nybble ptr + l;
end 2~ ;

Range Testing (RANGE) _ _ FIELD _ POS LEN DESCRIPTION
OPCODE 31: 28 4 constant: 4'bO000 _ _ _ _ _ , . _ _ TYPE 27 1 testing type: LTE (type = l'bO), GTE (type = l'bl) __ _ _ _ FIRST 261 nybble is first of compare word- l'bl, else first = l'bO
_ _ _ _ __ _ _ _ . _ __ _ VALUE _ 25: 22 4 compare value NEXT 21:00 22 next node address for successful ~omparison I _ ~

DESCRIPTION:
This is the range test instruction. It determines of a field in the header is either greater-than-or-equal-to or less-than-or-equal-to the compare value VALUE. "Big-endian" values are assumed when comparing strings greater than one nybble in length.
A satisfied bit, maintained by the Recognizer, is used to reminder that the inequality of strictly greater/less than has been achieved, and thus no more comparisons are necessary on remaining nybbles. The FIRST field should be set (TRUE) only duriny comparison of the first nybble in the string of interest.
If the compare succeeds, control is transferred to NEXT, otherwise, process-compare~failure is called for further processing. Successful comparison advances nybble_ptr by 1.

COMPUTATION:
instrl = RANGE .instruction format;
if (instrl[FIRST] == TRUE) begin satisfied = FALSE;
end - `

,C~ ~ 2 ~ Q ~

if (!satisfied) if (nybble == instrl [VALUE]) pgm adr = instrl[NEXT];
else if (instrl[TYPE] -= l'bl) // test for ~ =
if (nybble > instrl[VALUE]) satisfiad = TRUE;
pgm_adr - instrl[NEXT];
end else process_compare_failure else // test for < =
if (nybble < VALUE) begin satisfied = TRUE;
pgm adr = instrl[NEXT];
end else process_compare_failure;
pgm_adr = instrl~NEXT];
nybble ptr = nybble ptr + l;

2 1 2 ~

Skip Nybbles (SKIP) _ _ ____ , _ _ __ . _ _ _ FIELD POS LEN DESCRIPTION
. ._ _ . _ _ __ OPCODE 31:28 4 constant~ 4'bO010 _ _ ~
NUM27:22 6 number of nybbles to skip (2's-complement) , _ _ _ _ . _ _ NEXT 21:00 22 next node address after 5kip _ . _ _ . _ __ . __ _ _ _ _ DESCRIPTION:
Skip NUM nybbles in the packet header . The num field is a 2's-complement value su~h that the skip range is: -32 to + 31. ~ote that "SKIP 0" is a no-op.

COMPUTATION:
instrl = SKIP instruction format;
nybble ptr = nybble ptr + instrl[NUM]; // 2's complement addition Jump (JMP) , _ _ ~_ - - _ . . _ ._ . - .
FIELD POS LEN DESCRIPTION
_ _ _ OPCODE 31:25 7 constant: 7'bO011000 __ _ _ NUM 24:22 3 constant: 3'bO00 NEXT21-00 22 next node address for transfer I . ~

DESCRIPTION:
Cause an unconditional transfer o~ control. Parsing continues at the node specified byn nxt mode. This operation does not alter nybble ptr.

'2. 1 2 6 ~

COMPUTATION:
instrl = JMP instruction format;
pgm adr = instrl[NEXT~;

Jump to Subroutine (JSR) . _ _ -- -- ~
FIELD POS LEN DESCRIPTION
_ _ _ _ _ OPCODE 31:25 7 constant: 7'bO011001 I . _ _ _ _ . _ __ _ _ _ .' ¦S OE 24 1 source compare enable TSE 23 1 timestamp enable __ _ __ SPARE 22 1 constant: 1'bO
__ _ _ _ _ _ _ _ NEXT 21:00 22 next node address for transfer _ _ _ = _ _ _ _ . _ _ DESCRIPTION:
Jump to subroutine. Parsing continues at the node specified by NEXT. The value of NEXr is saved in srr. The SCE f ield enables Source Compare Mode and TSE enables Timestamp Mode. Only a single level of subroutine is supported. This operation does not alter nybble ptr.

COMPUTATION:
instrl = JSR instruction format;
mcm = TRUE; // enable modified compare mode mcm Ptr = nybble ptr; // save in case of compare failure if (instrl[SCE] == ~rRUE) // source compare mode enabled scm = l;

2 1 ?~

if (instrl[TSE] == TRUE) // timestamp mode enabled tsm = 1 srr = instrl[NEXT];
pgm_adr = instrl[NEXT~;

Case (CASE~
_ _ - ~
FIELD ROS LEN DESCRIPTION
_ _ _ _ _ _ _ OPCODE31:29 7 constant: 3'bllO _ _ _ _ _ SPARE28:22 3 _ constant: 7'bO000000 ' NEXT21-00 22 next node address for transfer ~ -- _ - ~

DESCRIPTION:
This performs a 16-way case branch by taking the next nybble from the header and logically ORing it with the low 4 bits of NEXT
field, thus effecting up to a 16-way branch. This instruction increments nybble ptr by 1.

COMPUTATION:
instrl = CASE instruc~ion format;
nxt_addr = ~18'bO, nybble] ¦ instrl~NEXT];
nybble_ptr = nybble ptr + 1;

-21~2 ~J',~ ~t, Return ~rom Subroutine (RTS) - . _ _ . _ _ . . , FIELD POS LEN DESCRIPTION
_ . . , _ OPCODE 31:25 4 cons~ant: 7'blO11011 I . _ _ . _ SPARE 241 constant: l'bO
_ _ _ ......... _ _ ___ TS 231 timestamp request _ ~ _ __ SPARE 22:21 2 constant: 2'bOO
j _ ._ ._ _ _ __ _ _ _ INDEX 20:04 17 index into Timestamp table (valid only when TS == 1) _. _ ___ _ OFFSET 3:0 4 return address offset . . .- ~

DES CRI PTI ON:
Return from subroutine with side-effect. The next node address in the upperl8 bits of srr concatenated with the 4-bit value in OFFSET. If TS is asserted, INDEX is loaded into tx dex. This operation increments nybble ptr by 1.

COMPUTATION:
instrl = RTS instruction format;
nybble_ptr = nybble ptr ~ 1:
next_addr = (srrt21:4], instrl[OFFSET]);?? explicit return if (instrl[TS] == TRE) if (tsm == TRUE) // is timestamping enabled?
ts dex = instrltINDEX]
ts valid = TRUE; // alert ALF SMC
end r~

2 ~ 2 ~

Done (DONE~
_ _ _, _ _ __ __ _ _ _ FIELD POS LEN DESCRIPTION_ _ _ OPCODE 31:2S 7constant: 7'b0011100 _ . _ _ _ __ _ SPARE 24-20 5 constant: 5'bOOOOO
. _ _ __ PID 19:00 20 packet Pattern ID
_ _ - - ~

DESCRIPTION:
Terminate header parsing operation. This packet's Pattern ID is in PID. The value of nybble ptr is no longer significant after executing this instruction.

COMPUTATION:
instrl = DONE instruction format;
pattern_id = instrl~PID];
nybble+ptr = unknown:
terminate:

Some of the instruction set pseudo-code described in this section uses a prcess compare_failure task to handle compare failures during compari~on instructions. That task is shown in Figure 4-1.

task process_compare failure;
i~ (mcm == TRUE) begin // modified compare mode mcm - FALSE; // reset for return pgm adr = ret adr[21:4] 1 4'bOOOl; // implicit return nybble_ptr = mcm=ptr]; // point to beginning of field nybble = fifo[nybble ptr]; // get next to nybble r~

2~26 3~ ~

if (scm == TRUE) begin // handle source compare mode scr ~tr = mcm ~tr; // save for ALF SMC
scr = TRUE; // alert ALF SMC
end end else begin // this is a termination fail gfc = TRUE; // set general failure flag terminate:
end endtask A parkicular implementation of the recognizer engine and memory controller is illustrated in Figure 5. This implementation is organized in three major logical units together with various latches and multiplexers. The three major logical units are an execution controller 61, a fetch controller 63 and an optimistic branch calculator 65. The execution controller and the fetch controller are implemented as finite state machines while the optimistic branch calculator is implemented as combinatorial logic.

As suggested by its name, the function of the fetch controller 63 is essentially to interface with the parse graph memory 57, presenting addresses to it and receiving back words of data representing parse graph instructions. The principal 2~2~5 ~

function of the execution controller 61 in turn is to analyze the instructions fetched, to read in corresponding nybbles from the FIFO, and to perform the comparisons called for by the various instructions.

While the execution unit generates definitive addresses resulting from the execution of instructions, the fetch controller initiates fetch operations based on an optimistic assumption, i.e~, that a needed comparison will succeed. The function of the optimistic branch calculator 65 is to determine such next instruction addresses proceeding on the assumption that the comparison will succeed.

Words read out of the parse graph memory are provided to the input designated by reference character 80, and are, in turn, provided directly to the optimistic branch calculator 65 and, through latches 67 and 69, to the execution controller 61. Two latches are provided since one instruction comprises two memory words as described previously. Since the most commonly encountered instruction types in the pars~ graph memory will be ones which merely designate a new node instruction address if a compare succeeds, the function of the optimistic branch calculator is merely to provide the 22 bits of the instruction to the memory address latch 71 as the next node address to be read.
Fetch addresses initiated in this fashion are directly coupled to the address register 71.

As indicated previously, the optimistic branch calculator 65 does not actually perform the compare which must succeed if the address is to be valid. Accordingly, provision for effectively ~lushing this read operation is implemented. The actual comparing is performed by the execution controller 61 which 2 ~ 2 ~

utilizes the last instruction read together with the corresponding nybble taken from the input FIFO. In other words, the execution controller 61 more completely analyzes the last read instruction in conjunction wlth a nybble being obtained from the input FIFO and correctly determines the appropriate action to be taken to execute the actual instruction. The address previously generated by the optimistic branch calculator 65 is provided back to the executive controller 61 through a latch 730 .

If the executive controller 61 determines that the memory read previously initiated by the optimistic branch calculator 65 is inappropriate, it can signal to the fetch controller that the previous operation should be flushed and can provide a flush address to the memory address latch 71 through a 3 to 1 multiplexer 77. Multiplexer 77 also allows the memory address latch 71 to be loaded with an address provided by the control processor 31 through the slave interface unit (SIU). The SIU
also allows the control processor 31 to reset the fetch controller as indicated at 79. In addition, acting through the slave interface unit, the host computer 17 and control processor 31 can load the parse graph memory 57 through gates 81 which can drive the memory data lines for a write operation.

The parse graph instructions in general comprise two different types. The first o~ these may be referred to a node instruction. The logic means or recognizer engine responds to such node instructions by initiating a memory read at a next address in the parse graph memory 57 which, given the state of the recognizer , is determinable ~rom only the current node instruction and the current header bit group. For each of these node instructions there is also a default condition which indicates that recognition has failed and, upon this condition, ~1.2~

the logic means 25 generates output signals or data which communicates that fact to the ALF-SMC. In effect this means that the message header is of a new type which has not previously been characterized and for which no ~orresiponding instruction pattern has been set up in the parse graph memory.

The second type of instruction is a termination irtstruction.
Encountering a termination instructiort irt the processing of ~he message header means that the header has been sttccess~ully classified or recognized so that no further processing is needed.
Each termination instruction includes a pattern ID which identifies the corresponding previously characterized message header and associates it with a path through the ATM network. As indicated previously, the ALF-SMC 38 utilizes this pattern ID to control the header management and segmentation circuitry 27 to convert the incoming local message into a sequence of ATM cells which can proceed through the ATM network and then be reassembled at the interface with the destination device or local area network. The controlling data provided by the logic means or classifier 25 to the AI.F-SMC may also simply tell that device to discard the message since one of the operations of the classifier may be to use instruction patterns in the parse graph memory 57 which cause messages to not be accepted into the ATM network, e~g., when functiolling as a bridge or filter as described previously.

Typically, the operation of the ALF-SMC in response to a recognition failure will be to signal to the control processor 31 and/or the host computer 17 to cause it to execute a more conventional program for analyzing the messaye header to determine if the message is one being of any interest of the ATM
system. If it is of interest, the control processor can then ~2~

write a corresponding pattern of instructions into the read/write memory 57 so that, from then on, the message header will be recognized or classified by the logic means 53 and memory 57 without involvement of the host computer. As will be understood from the foragoing description, the ~dded instructions will end with a termination instruction incorporating an appropriate pattern ID to enable the ALF-SMC to either appropriately convert or discard the message.

As will be understood by those skilled in the art a session of communications between a given source and destination will typically involve a large number of messages or packets.
Accordingly, the great majority of messages which must be considered will be those previously characterized in the sense that corresponding sequences of instruction have been written into the memory 57. Thus, only new header types will have to incur the time penalty of a fully general analysis by the host computer 17. Further, since there is a correspondence between the number of instructions which need to be executed with the length of the message header, it will be understood that the maximum time required to recognize or classify a header is roughly proportional to the length of a header and thus also the time required to present it sequentially.

In addition to using the pattern ID information to control header conversion and segmentation, the ALF-SMC 38 also provides this data to the control processor 31. The control processor 31 utilizes this i.nformation to maintain a table of how recently each type of message header has appeared. Using this information the control processor can then remove from the read/write memory 57 those sequences of instructions which are no longer needed, thus ~reeing up memory for the introduction of new instruction 2~ 2~

sequences when it is necessary to accommodate new message headers. In general it may be noted that the classifying scheme implemented by the present invention is relatively undemanding of processor speed but consumes a relatively large amount of memory.

As indicated previously, some common testing procedures are advantageously conducted using a form of subroutine in the parse graph. Subroutines can be entered by a Jump to Subroutine (JSR) instruction, a definition of which was given earlier. As will be noted from the explicit definition of the JSR instruction certain information regarding the state of the recogniæer logic is stored on execution of this instruction, which information is utilized in returning from the subroutine. Thus in the case of subroutines, the calculation of the next address in the parse graph memory may depend upon the stored state of the classifier logic as we all upon the current instruction and the nybble being read from the FIFO register. Return from subroutine is accomplished by the RTS instruction, also explicitly defined previously. The RTS instruction is somewhat unusual in that is allows the return to occur with an address offset. Accordingly, a subroutine can be structured with several RTS instructions with dif~erent o~fsets so that the point of return can be varied as a ~unction of determinations made during the subroutine.

As will be understood by those skilled in the art, it will not always be appropriate to terminate parsing and discard and incoming message just because the header is not completely recognized. For example, if the destination can be recognized, it may be appropriate to forward the message through the network while at the same time alerting the control computer and/or the host computer that the message originated from an unrecognized source. The Jump to Subroutine (JSR) instruction implements this capability by putting the recognizer into a modified compare mode by setting the mcm bit. When this bit is set, a comparison failure does not automatically initiate a termination of parsing and 2 signaling o~ the ALF-SMC that a failure has occurred.
]Rather, parsing is continued and the decision to signal the ALF~
SMC is deferred until the return from the subroutine.

The JSR instruction also provides for entering a mode known as source compare mode. When the source compare enable (sce) bit is set in the JSR instruction, the results of the modified compare mode are saved for later use by the ALC-SMC.
Specifically, if a compare failure occurs when the modified compare mode is in effect, i.e., during a subroutine, the pointer (mcm_ptr) which identifies the instant header nybble is loaded into the source compare failure pointer (scf ptr) and the source compare failure status bit SCF is asserted. Thus, at the end of parsing, the ALF-SMC can he informed of the particular nybble within the header which caused the source compare failure. As will be understood by those skilled in the art, the inclusion of these subroutine facilities and in particular the implementation o~ modified compare mode and source compare mode add flexibility to a system which is inherently quite fast as described previously.

In view of the foregoing it may be seen that several objects of the present invention are achieved and other advantageous results have been attained.

As various changes could be made in the above construction without departing from the scope of the invention, it should be understood that all matter contained in the above description or shown in the accompanying drawings shall be interpreted as .

~:, , ' ,, . , ' ',; ,, ' . .' . , ', ., ' ' ' ' ' ' .: , : :: , . :

_34_ 21 26~j~3r~

illustrative and not in a limiting sense.

Claims (20)

What is claimed is:

1. In a device for directing selected messages to a selected one of a plurality of output paths in accordance with the contents of a message header which comprises a sequence of bit groups presented successively, apparatus for quickly classifying headers; said apparatus comprising:

memory for storing, at a multiplicity of addresses, node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for different header types;

logic means responsive to a node instruction read from memory for initiating a memory read at a next address determinable from the current node instruction and a current header bit group, said logic means being responsive to a terminator instruction for outputting respective header identifying data whereby, for each header type, a corresponding pattern of node instructions and a terminator instruction can be written into said memory, thereby enabling said apparatus to identify the respective header.

2. Apparatus as set forth in claim 1 wherein neither said node instruction nor said terminator instructions cause writing of data into said memory.

3. Apparatus as set forth in claim 1 wherein said logic means does not write data to said memory.

4. In a device for directing selected messages to a selected one of a plurality of output paths in accordance with the contents of a message header which comprises a sequence of bit groups presented successively, apparatus for quickly classifying headers; said apparatus comprising:

memory for storing, at a multiplicity of addresses, node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for different header types;

logic means responsive to a node instruction read from memory for initiating a memory read at a next address which, given the current state of the logic means, is determinable from only the current node instruction and a current header bit group, said logic means being responsive to a terminator instruction for outputting respective header identifying data whereby, for each header type, a corresponding pattern of node instructions and a terminator instruction can be written into said memory, thereby enabling said apparatus to identify the respective header.

5. Apparatus as set forth in claim 4 wherein said node instructions include a jump to subroutine instruction which causes a memory read at a next address determinable from the instruction and wherein said logic means includes a register for retaining data identifying the address of the jump to subroutine instruction.

6. Apparatus as set forth in claim 5 wherein said node instructions include a return from subroutine instruction which causes a memory read at a next address which is offset from the address of the jump to subroutine instruction by a preselectable offset.

7. In a packet switched device for directing selected digital messages to a selected one of a plurality of output paths in accordance with the contents of a message header which comprises a sequence of bit groups presented successively, apparatus for quickly classifying previously characterized headers; said apparatus comprising:

memory for storing, at a multiplicity of addresses, node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for previously characterized header types;

logic means responsive to a node instruction read from memory for initiating a memory read at a next address which, given the current state of the logic means, is determinable from only the current node instruction and a current header bit group or for outputting data indicating recognition failure if no next address is determinable, said logic means being responsive to a terminator instruction for outputting the respective header identifying data whereby, for each previously characterized header type, a corresponding pattern of node instructions and a terminator instruction can be written into said memory, thereby enabling said apparatus to identify the respective header.

8. Apparatus as set forth in claim 4 wherein said node instructions include a jump to subroutine instruction which causes a memory read at a next address determinable from the instruction and wherein said logic means includes a register for retaining data identifying the address of the jump to subroutine instruction.

9. Apparatus as set forth in claim 8 wherein said jump to subroutine instruction inhibits output of recognition failure data.

10. Apparatus as set forth in claim 8 wherein said node instructions include a return from subroutine instruction which causes a memory read at a next address which is offset from the address of the jump to subroutine instruction by a preselectable offset.

11. Apparatus as set forth in claim 7 wherein said node instructions include a compare for equality instruction which first masks a header bit group with a preselected mask value and generates a next instruction address if a match is obtained with a preselected compare value.

12. Apparatus as set forth in claim 7 wherein said node instructions include a multiple compare instruction in which a succession of header bit groups are compared with respective mask values and a next instruction address is generated if all matches are obtained.

13. Apparatus as set forth in claim 7 wherein said node instructions include a range instruction which generates a next instruction address if a header bit group falls within a given range of values.

14. Apparatus as set forth in claim 7 wherein said node instructions includes a case instruction which generates a next instruction address which is a function of the value of the current header bit group.

15. Apparatus as set forth in claim 14 wherein said case instruction includes a value which is logically ORed with the current header bit group.

16. In a device for selecting digital messages for further processing in accordance with the contents of a message header which comprises a sequence of bit groups presented successively, apparatus for quickly classifying previously characterized headers; said apparatus comprising:

a register for storing a plurality of said bit groups;

memory for storing, at a multiplicity of addresses, node instructions which comprise opcodes in association with respective next address characterizing data and terminator instructions which comprise identifying data for previously characterized header types;

logic means which reads bit groups from said register and instructions from said memory and is responsive to a node instruction read from memory for initiating a memory read at a next address which, given the current state of the logic means, is determinable from only the current node instruction and the current header bit group, said logic means being responsive to a terminator instruction for outputting the respective header identifying data whereby, for each header type, a corresponding pattern of node instructions and a terminator instruction can be written into said memory, thereby enabling said apparatus to output the respective header identifying data.

17. Apparatus as set forth in claim 16 wherein said node instructions include a skip instruction which cause said logic means to pass by a selected number of said bit groups in reading from said register.

18. Apparatus as set forth in claim 16 wherein said node instructions include jump instructions which generate next address values independently of bit group values.

19. Apparatus as set forth in claim 16 wherein neither said node instruction nor said terminator instructions cause writing of data into said memory.

20. Apparatus as set forth in claim 16 wherein said logic means does not write data to said memory.