CA1136728A - Multiprocessor system - Google Patents

Abstract

Abstract of the Disclosure A multiprocessor system the kind in which two or more separate processor modules are interconnected for parallel processing includes two redundant interprocessor buses dedicated exclusively to interprocessor communication. Any processor module may send information to any other processor module by either bus. The buses are shared in use by the processor modules on a time-sharing basis. Use of each bus is controlled by a special bus controller. The multiprocessor system includes an input/output system having multi-port device controllers and input/output buses connecting each device controller for access by the input/output channels of at least two different processor modules. Each device controller includes logic which insures that only one port is selected for access at a time. The multi-processor system includes a distributed power supply system which insures uninterrupted operation of the remainder of the multi-processor system in the event of a failure of a power supply for a part of the system. The distributed power supply system includes a separate power supply for each processor module and two separate power supplies for each device controller. Either one of the two power supplies provides the entire power for the device controller in the event the other power supply fails.
The distributed power supply system permits any processor module or device controller to be powered down so that on-line mainten-ance can be performed in a power-off condition while the rest of the multiprocessor system is on-line and functional. The multiprocessor system includes a memory system in which the memory of each processor module is divided into four logical address areas -- user data, system data, user code and system code. The memory system includes a map which translates logical addresses to physical addresses and which coacts with the multiprocessor system to bring pages from secondary memory into primary main memory as required to implement a virtual memory system. The map also provides a protection function. It provides inherent protection among users in a multiprogramming environment, isolates programs from data and protects system programs from the actions of user programs. The map also provides a reference history information for each logical page as an aid to efficient memory management by the operating system.
The multiprocessor system includes in the memory of each processor module an error detection and correction system which detects all single bit and double bit errors and which corrects all single bit errors in semiconductor memory storage.

Description

~13bf~P~

This applica-tion is a divisional of ~opending Canadian Patent Appllca-tion Serial Mo. 280,505 ~iled June 14, 1977 in the name of Tandem Compu-ters Incorporated.
This invention relates to a multiprocessor com-puter system in which interconnected processor modules provide multirrocessing (parallel processing in separate processor modules)and multiprogramming (interleaved pro-cessing in one processor module).
This invention relates particularly to a system which can support high transaction rates to large on-line data bases and in which no single component failure can stop or contaminate the operation of the system.
There are many applications which require on-line processing o~ large volumes of data at high trans-action rates. For example, such processing is re~uired in retail applications for automated point of sale, inventory and credit transactions and in financial institutions for automated funds transfer and credit transactio~s.
In computing applications of this kind it is important, and often critical, that the data processing i not be interrupted. A failure of an on-line computer system can shut down a portion of the related business and can cause considerable loss of data and money.
` Thus,-an on-line system o~ this kind must provide not only suEficient computing power to permit multiple computations to be done simultaneously, but it must also provide a mode o~ opexation which pexmits data processing to be continued without interruption in the event some com-ponent of the system fails, The system should operate either in a fail-safe mode (in which no loss of throughput occurs as a result of I

1~3tj7Z~

1 failure) or in a fail~soft mode (in ~Jhich some slo~7down

2 occurs but full processing capabilities are maintained)

3 in the event af a failure.

4 Furthermore, the system should also operate in

5 a way such that a failure of a single component cannot

6 contaminate the operation of the system. The system should

7 provide fault-tolerant computing. For fault-tolerant

8 computing all errors and failures in the system should either

9 be corrected automatically, or if the failure or error cannot

10 be corrected automatically, it should be detected, or i~ it

11 cannot be detected, it should be contained and should not

12 be permit~ed to contaminate the rest of the system.

13 Since a single processor module can fail, it is 1~ obvious that a system which will operate without interruption in an on-line application must have more than one processor 16 modale.
17 Systems which have more than one processor module 18 can therefore meet one of the necessary conditions for non-19 interruptible operation. However, the use of more than one 20 processor module in a system does not by itself provide all 21 the sufficient conditions for maintaining the required 2 processing capabilities in the e~ent or component ailure, 23 as will become more apparent from the description to follow.
24 Computing systems for on-line, high volume, trans-;... . .
25 action oriented, computing applications which must operate 2~ without interruption therefore require multiprocessors as a 27 starting point. But the use of multiprocessors does not 28 guarantee that all of the sufficient conditions will be met, 29 and fulfilling the additional suf~icient conditions for on-line ,~ .

~13tj 7Z~

1 systems of this kind has pres~nted a ~umh~r of probl~ms 2 in the prior art.
3 The prior art approach to uninterruptcd data 4 processing has proceeded generally along two lines - either 5 adapting t~o or more large, monolithic, general purpose 6 computers for joint operation or interconnecting a plurality 7 of minicomputers ~o provide multiprocessing capabllities.
8 In the first case, adapting two large monolithic g general purpose computers for joint operation, one conven-10 tional prior art approach has been to have the two computers 11 share a common memory. Now in this type of multiprocessing 12 system a failure in the shared memory can stop the entire 13 system. Shared memory also presents a number of other

14 problems includingsequencing accesses to the common memory.

15 This system, while meeting some of the necessary conditions

16 for uninterruptible processing, does not meet all of the 1;7 sufficient conditions.
18 Furthermore, multiprocessing systems using large 19 general purpose computers are quite expensive because each 20 computer is constructed as a monolithic unit in which all 21 components ~including the packaging, the cooling system, 22 etc.) must be duplicated each time another processor is 23 added to the system even though many of the duplicated 24 components are not re~uired.
The other prior art approach of using a plurality of 26 minicomputers has (in common with the approach of using large 27 general purpose computers) suff~red from the drawback of 28 having to adapt a communications link between computers 29 that were never originally constructed to provide such a 30 link. ~he required links were, as a result, usually made ~3u72~

1 through the input/output channel. Connections through the 2 input/output channel are necessarily slower than internal 3 transfers within the processor itself, and such interproCessOr 4 links have therefore provi~ed relatively slow interprocessor S communlcation.
6 Furthermore, the interprocessor connections 7 required special zdapter cards that added substantially to 8 the cost of the overall system and that introduced the g possibility of single component failures which could stop 10 the system. Adding dual interprocessor links and adapter 11 cards to avoid problems of critical single components failures 12 increased the overall system cost even more substantially.
13 Providing dual links and adapter cards between 14 all processors generally became very cumbersome and quite 15 complex from the standpoint of operation.
16 Another problem of the prior art arose out of the

17 way in which connections were made to peripheral devices.

18 ~ If a number of peripheral devices are connected to l9 a single input/output bus of one processor in a multiprocessor 20 system and that processor fails, then the peripheral devices 21 will be unavailable to the system even though the failed 22 processor is linked through an interprocessor connection to ~3 another processor or processors in the system.
24 To avoid this problem, the prior art has provided an 2S input/output bus switch for interconnecting input/output busses 26 for continued access to peripheral devices when a processor 27 associated with the peripheral deviccs on a particular input/
28 output bus fails. The bus switches have been expensive and also 29 have presented the possibility of single component failure 30 which could down a substantial part of the overall system.

~ ~Jt~2 ~

1 Providin~ softwarc for the p~ior ~rt m~ltiproc~ssor 2 systems has also been a major problem.
3 Operating systems software for such multiprocessing 4 systems has tended to be nonexistent. Where software had been devel~ped for such multiprocessor systems, it quite 6 often was restricted to a small number o~ processors and 7 was not adapted for the inclusion of additional processors.
8 In many cases it was necessary either to modify the operating 9 system or to put some of the operating system functions into the user's own program -- an expensivel time-consuming 11 operation.
12 The prior art lacked a satisfactory standard operating 13 system for linking processors. It also did not provide an 14 operating system for automatically acco~nodating additional processors in a multiprocessing system constructed to 16 accommodate the modular addition of processors as increased 17 computering power was required.
18 A primary object of the present inven~ion is to

19 construct a multiprocessor system for on-line, transaction-

20 oriented applications which overcomes the problems of the

21 prior art.

22 A basic objective of the present invention is to ~3 insure that no single failure can stop the system or significantly 24 affect system operation. ~n this regard, the system of 25 the present invention is constructed so that there is no 26 single component that attaches to everything in the system, 27 either mechanically or electrically.
28 It is a closely reLated objective of the present 29 invention to guarantee that every error that happens can be 30 either corrected, detected or prevented from contaminating the syste ~13~'72E~

1 It is another important objective of th~ present 2 invention to provide a s~stem architecture and basic mode 3 of operation which free the uscr from the need to c3et involved 4 with the system hardware and the protocol of interprocessor communication. In the present invention every major component 6 is modularized so that any major component can be removed or 7 replaced without stopping the system. In addition, the 8 system can be expanded in place (either horizontally by the g addition of standard processor modules or in most cases vertically by the addition of peripheral devices) without ll system interruption or modification to hardware or software.

.15

23

24 ~3~Z~

1 Summ~y of th~ InventiGn , 3 The multiprocessor system of the present invention 4 comprises multiple, independent processor modules and data 5 paths.
6 In one specific embodiment of the present invention 7 16 separate processor modules are interconnected by an 8 interprocessor bus for multiproces5ing and multiprogramming.
In this specific embodimen~ each processor module supports ~0 up to 32 device controllers, and each device controller can 11 control up to eight peripheral devices.
12 Multiple~ independent communication paths and ports 13 are provided between all major components of the system to 14 insure that it is always possible to communicate between 15 processor modules and between processor modules and peripheral 16 devices over at least two paths and also to insure that a 17 single failure will not stop system operation.
18 ~ These multiple communication paths include multiple 19 interprocessor busses interconnecting each of the processor 20 modules, multiports in each device controller, and input/output 21 busses connecting each device controller for access by at 22 least two different processor modules.
23 Each processor module is a standard module and 24 includes as part of the module a central processing unit, a

25 main memory, an interprocessor control and an input/output

26 channel.

27 Each processor moduLe has a pipelined microprocessor ~8 operated by microinstructions included as a basic instruction 29 set in each processor module.
The basic instruction set in each ~367~:~

1 processor module recogni~cs the fact that there is an 2 interprocessor communications link; and ~rhen an ad~itional 3 processor module is added to the system, the operating system 4 ta copy of which resides in each processor module) is informed 5 that a new resource is available for operation within the 6 existing operating system without the need to modify either 7 the system hardware or software.
8 To increase performance and to maintain very high 9 transaction rates each processor module includes a second 10 microprocessor which is dedicated to input/output operations, 11 A dual port access to the main memory by both the 12 central processing unit and the input/output channel permits 13 direct memory access for the input/output transfers to also 14 increase performance.
Each processor module is physically constructed 16 to fit on a minimum number of large printed circuit boards.
~7 Using only a few boards for each processor module conserves 18 space for packaging and minimizes the length of the inter-l9 processor bus required to interconnect all of the processor 20 modules. A relatively short interprocessor bus minimizes the 21 deterioration of the signals on the interprocessor bus and 1 22 permits high speed of communication over the interprocessor 23 bus.
24 Each interprocessor bus is a hiyh speedj synchronous 25 bus to minimize o~erhead in interprocessor communications and 26 to enable the system to achieve high throughput rates.
27 A separate bus controller monitors all transmissions

28 over the bus. The bus controller includes processor select

29 logic for determining the priority of data transfer between

30 any two processor modules over the interprocessor bus. The ~13~ 72~

1 bus controller also lncludes bus control state lo~ic for 2 establishin~ a sender-receiver pair of processor modules 3 and a time frame for a transfer of information over the bus 4 between the sender-receiver pair.
Each bus controller includes a bus clock, and each 6 central processing unit of each processor module has its o~m 7 separate cloek. There is no master clock system subject to 8 a single component failure which could stop the entire 9 multiprocessor.system.
Each processor module includes, in the interprocessor 11 control o~ the processor module, a certain amount of circui~ry 12 on the printed circuit boards which is dedicated to communications 13 over the interprocessor buses, 14 Each interprocessor control also includes fast ~5 buffers (inqueue buffers and an outqueue buffer) which can 16 be emptied and filled by the central processing unit withou~
17 interfering with the interprocessor bus. This makes it 18 possible to sustain a higher data rate on the interprocessor 19 bus than could he sustained by any single pair o~ processors.
Several data transfers between pairs of processor modules 21 can be interlea~ed on an apparent simultaneous basis.
22 Because the interprocessor bus operates asynchronously 23 with each particular central processing unit, eaah inqueue 24 and outqueue buffer is clocked either by the processor module or by the bus controller, but not by both simultaneously.
26 Each inqueue buffer and outqueue buf~er therefore 27 has associated with it in the interprocessor control some 28 logic that operates in synchronism with the bus clock and ~9 other logic that operates in synchronism with the central 3~ processing unit clock. Logic interlocks qualify certain 1 transitions of the lo~ic from one state to another state 2 to prevent loss of dat~ ln transfers bet-~een the asynchronous 3 interprocessor huses and processor module.
4 The logic is also arranged so that in the event S a processor module is powering down, there will be no transi~nt 6 effect on the inte-processor buses because the processor module 7 is losing control. The powering do~m of the proce~sor ~odule 8 on an interprocessor bus will therefore not disrupt any ot~er g interprocessor bus activity.
The bus controller and interprocessor control of 11 each processor module coact to perform all interprocessor 12 bus management in parallel with processing by the central 13 processing units so that there is no waste of processing 14 power. This bus management is performed with low lS protocol overhead in that it takes very few interprocessor 16 bus cycles ~o establish a bus transfer -- what processor 17 bus module is sending and what processor module is receiving --18 relative to the amount of information actually transmitted.
19 The processor select logic of the bus controller 20 includes an individual select line which extends from the 21 processor select logic to each processor module~ The select 22 lines are used in three ways in the protocol of establishing 23 a sender-receiver pair of processor modules and a time 24 frame for transfer of information over the interprocessor 25 bus between the sender-receiver pair. The seleat lines are 26 used (1) in polling to determine ~Ihich particular processor 27 module wants to send, (2) in receiving to inquire of a receiver 28 processor module whether the particular processor module wants to 29 receive, and (3) in combination with a send command to let the 30 sender processor module know the time ~rame for sending.

i7;2~

1 The r~eeiver proc~sso~ module is qualified to 2 receive incoming data unsolicited by the receiver processor 3 module and without a software instruction.
4 Blocks of data bet~een a sender-receiver pair of processor modules are transmitted over the interprocessor 6 bus in packets. At the end of each packet transfer the 7 interprocessor control of a receiver processor module logically 8 disconnects from the interprocessor bus to permit the bus g control state logic to establish another sequence of a lO different sender~receiver pair of processor modules and a 11 time frame for making a packet trans~er between the other 12 pair of sender-receiver processor modules. Thus, as noted 13 above, several data block transfers between diferent sender-14 receiver pairs of processor modules can therefore be interleaved 15 on the interprocessor bus on an apparently simultaneous basis 16 because of the faster clock rate of the interprocessor bus as 17 compared to the slower memory speed of the processor modules.
18 Each processor module memory includes a separate 19 buffer for each combination of a processor module and an 20 interprocessor bus.
21 Each memory also includes a bus receive table for 22 directing incoming data from an interprocessor bus to a 23 specified locakion in a relàted buffer in the memory of a 24 receiver processor module. Each bus receive table provides 25 a bus receive table entry which contains the address where the 26 incoming data is to be stored and the number of words expected 27 from the sender processor module. The bus receive table 28 entry is updated by firmware in the processor module after 29 the receipt of each packet and i5 effective with the firmware 30 either to provide a program interrupt when the entire data 1~

1~31~72B

1 block has been successfull~ rec~ived or to provid~ an interrupt 2 to the softwaxe progr~m currently executirly in the processor - 3 module in response to the detection of an error in the course 4 of the transmission of the data over the interprocessor bus.
Producing a program interrupt only at the completion of the 6 data block transfe~ enables the transfer of data to be made 7 transparent to the software currentl~ executing in the 8 processor module. The interrupt in response to the detection 9 of an error provides an integrity check on the transmission of data.
11 The input/output subsystem of the multlprocessor 12 system of the present inven~ion is constructed to insure that 13 no single processor module failure can impair system operation.
14 In addition, the input/output subsystem is 15 constructed to handle very high transaction rates, to 16 maximize throughput, and to minimize interference with 17 programs running in the processor modules.
18 ~ As noted above, each processor module includes a 19 microprocessor which is dedicated to input/output operations~
The input/output system is an interrupt driven 21 system and provides a program interrupt only upon completion 22 of the data transfer. This relieves the central processing 23 unit from being dedicated to the device while it is transferring 24 data.
Each input/output channel is block multiplexed to 26 handle sever~l block transfers o~ data ~rom several device 27 controllers on an apparent simultaneous basis. This is 28 accomplished by interleaving variable length bursts of data 29 in transfers between the input/output channel and stress 30 responsive buffers in the device controllers.

~i7~

1 As noted above, each device controller has multiports, and a separate input/output bus is connected to each port 3 so that each device controller is connected for acc~ss by 4 at least two different processor mcdules.
The ports of each device controller are constructed 6 so that each port is lo~ically and physicall~ independ~nt 7 of each other port. No component part of one port is also a 8 component of another port so that no single component failure 9 in one port can affect the operation of another port.
Each device controller includes logic which insures ~ that only one port is selected for access at a time so that 12 transmitting erroneous data to one port can never contaminate 13 another port.
14 The input/output system of the present invention lS interfaces the peripheral devices in a failsoft manner. There 16 are multiple paths to each particular device in case of a 17 failure on one path. And a failure of the device or a failure 18 of a processor module along one path does not affect the 19 operation of a processor module on another path ~o the device.
The input/output system of the present invention 21 is also constructed so that an~ type of device can be put 22 on the system, and the input/output system will still make 23 maximum usage of the input/output channcl bandwidth.
24 The device controllers are buered such that all 25 trans~ers between the device controllers and the input/output ~6 channel occur at the maximum channel rate.
27 The device controller may transfer between itself 28 and a peripheral de~ice in bytes, but the device controller 29 must pack and unpack data to tr~nsfer words between itself 30 and the input/output channel.

13 _, ~,1~;7~

1 Because the buffers are located in thc device 2 controllers rather than in the input/output channel, the 3 present invention limits the buffering to onl~ the buffering 4 required by a particular system configuration. The present invention does not require a separate buffer for each peripheral 6 device in order to prevent overruns, as would be required 7 if the buffers were located in the input/output channel 8 rather than in the device controllers as had often been the 9 practice in the prior art.
As noted above, each buffer is a stress responsive 11 buffer and this provides two advantages.
12 First of all, each buffer can be constructed to 13 have an overall depth which is related to the type and number 14 Of devices to be serviced. Each device controller can therefore 15 have a buffer si~e which is related to the kind o~ devices 16 to be controlled.
17 Secondly, the stress responsive buffer construction 18 and mode of operation of the present invention allows the 19 buffers to cooperate without communicating with each other.
20 This in turn permits optimum efficient use of the bandwidth 21 of the input/output channel.
22 The stress placed on a particular buffer is determined 23 by the degree of the full or empty condition of the bufer 24 in combination ~7ith the direction of the transfer with respect 25 to the processor module. Stress increases as the peripheral 26 device accesses the buffer, and stres~ decreases as the input/
27 output channel means access the buffer.
28 Each buffer has a depth which is the sum of a 29 threshold depth and a holdo~ depth. The threshold depth 30 is related to the time required to service higher priority Z~

device controllers, and the holdoff depth is related to the time required to service lower priority device controllers connected to the same input/output channel.
The stress responsive buffer includes control logic for keeping track of the stress placed on the buffer.
The control logic is effective to make reconnect requests to the input/output channel as the stress passes through a threshold depth of the buffer.
Each buffer having a reconnect request pending is individually connected to the input/output channel in accordance with a polling scheme which resolves priority among all the device controllers having a reconnect request pending.
When the device controller is connected to the input/output channel, the data is transferred between the buffer and the input/output channel in a burst at or near memory speed.
Thus, because the buffers transmit data to and from the peripheral devices at the relatively slow device speed and can transmit the data to and from the processor modules at or near memory speed in burst transfers, and in response to buffer stress, the burst transfers can be time division multiplexed so that individual bursts from several device controllers can be interleaved to optimize efficient use of the bandwidth of the input/output channel and also to permit several block tran8fers from different device con-trollers to be made on an apparent simultaneous basis.
Comprehensive error checks and provision for error containment are provided for all data transfers over the data paths of the multiprocessor system.

.

1~367~

The error checks include check su~ning and par-ity checks on the da-ta paths and error detection and correction in the main memory system.
The error checks also include time ou-t llmita-tions in the input/output channel.
Error containment is provided in the input/
output system by an input/output control table having a two-word entry for each peripheral device to define a buffer area in the memory ~or the particular device controller and device. Each two-word entry describes the buffer location in main mernory and the remaining byte count length to be transferred at any particular time for a particular data transfer to a device. The input/output control table is located in each processor instead of in the device controllers to contain the results of any failure in the countword or address word to the single processor module in which the countword or address word is physically located. Each of the processor modules that is connected for access to common device controllers and related devices contains its own copy of the input/output control table. The failure of a table entry in one pro-cessor module does not affect the other processor module because the other processor module has its own correct copy of the ta~le entry.
The multiprocessor system of the present inven~
tion includes a power supply system which distributes separate power supplies ko the processor modules and device controller9 in a way ko insure unintcrrupted operation of the remainder of the multiprocessor system in the event of failure of a power suppl~ or part of the multiprocessor system.

31 1367Z8 1 Any processor mo~ule or device controller can be 2 powered down so that on-line maintenance can be perforrned in 3 a powered-off condition while the rest of the multiprocessor 4 system is on-line and functional.
The power supply system includes a separa~e po~ter 6 supply for each processor module and t~.ro separate power 7 supplies ~or each device controller.
8 The two separate power supplies are operatively 9 associated with the device controller by a switch which permits one power supply to supply all of the power for the ll device controller in the event of a failure of the other 12 power Supply, 13 The power supply system of the present invention 14 also produces a power failure warning signal which is effective to save the state of the logic in a processor module 16 in the event of a ~ailure of a power supply associated with 17 that processor module. When power is restored, the processor 18 module is returned to operation in a state that is known and l9 without the loss of data.
The memory of the multiprocessor system of the 21 present invention is divided into four logical areas -- user 22 data, system data, user code and system code. This division 23 Of memory into four separate logical address areas separates 24 code from data so that code can be made nonmodifiable and also separates operating system programs from user programs 26 so that users cannot inadvertently destroy the operating 27 system.
28 The multiprocessor system of the present invention 29 includes a memory map which perfoxms a number of functions.

~72~il 1 One function of th~ map is to provide a virtual 2 memory system in which all code and data are inherently 3 relocatable so that the user need not be concerned with the 4 actual physical location of either system or user programs 5 or the amount of physical memory a-ttached to the system.

6 The map translates logical addresses to physical 7 addresses for pages in main memory and provides paye fault 8 interrupts for pages not in main memory. The operating g system brings pages from secondary memory (i.e., memory 10 stored in peripheral devices) into the primary main memory 1~ in the processor module as required to implement a virtual 12 memory system in which the physical page addresses are invisible 13 to users and in which logical pages need not reside in 14 contiguous physical pages and need not be in physical main 15 memory but may be in secondary memory.

16 The map also provides a protection function and 17 a memory management function.

18 ~ The map provides a separate map for each separate 19 logical area of memory.

This provides protection by separating code from 21 data and also by separating the user programs from the 22 system programs, as pointed above.

23 It also provides protection among users in a 24 multiprogramming environment because the map which is in 25 effect for a particular user points only to the physical 26 memory pages of that user's program. This prevents one 27 user from writing into a program page of another user's 28 program. This feature of a user map therefore protects, 29 without the need or protection rec3isters, one user from 30 destroying another user's program.

1~

~13~ 8 1 The map in conjunction with the opera~ing syst~m 2 performs a map memory management function to r~duc~ opcr~tinq 3 overhead in the management of the memory system by (1) makiny 4 pages available from secondary memory, (2) keeping track S of frequency of use of physical pages in primary memor~, 6 (3) reducing virt~al memory page input/output transfers, 7 and (4) reducing-interrupts to the operating sys~em~ The 8 way that the map accomplishes these functions provides g an efficient virtual memory system.
iO The number of pages available in physical main 11 memory is limited. Physical pages must therefore sometimes 12 be brought into physical main memory from secondary memory.
13 One important aspect of efficient memory management 14 is to keep track of what pages in physical main memory are 15 being used frequently enough so as to need to be retained 16 in physical main memory.
17 Another important aspect is to know whether any 18 particular pages in physical main memory can be written 19 over (overlaid) without having to be first swapped out to 20 secondary storagP.
21 The map includes history bits as a part of the-map 22 entry for each page. These history bits (which are physically 23 in the map entry) give a histogram of usage of the given 24 physical page over a period of time. And, in the present 25 invention, the history bits are periodically updated by 26 hardware without the need for program intervention.
27 Each map entry also includes a "dirty bit" for 28 indicating whether a particular page has been written into 29 since it was last brought in ~rom secondary storage.

lg ~36~

1 The map therefore includes in thc map itsel~
~ information which permits the memory manager to determine 3 whether a particular page in physical main memory is a good 4 candidate for being overlaid (~"hen it i~ necessary to bring a page in from secondary storaye and no empty page or code 6 page in physical m~in memory is available for an overlay) 7 and to determine also, if an overlay is required, whether 8 or not it is necessary to swap the overlaid page out to, g seeondary storage before the page can be overlaid. Since eopies of all non-dirty pages are kept in secondary storage, 11 -no swap is required if the dirty bit is not on.
12 The map is eontained in a part of the memory which 13 is separate from the main memory. Each map is eonstructed to 14 provide significantly faster access than the aceess to physical main memory so that the map ean be rewritten in the 16 time that a physical memory aeeess is being aceomplished.
17 The rewriting of the map therefore does not increase memory 18 Cyele ~ime, 19 As noted above, the memory includes dual port access for the eentral processing unit and the input/output ehannel.
2L The input/output channel can therefore aeeess the memory 22 direetly, without having to go through the central proeessing 23 unit, for data transfers to and from a devlee controller.
24 Central processing unit accesses to memory and input/output ehannel aeeesses to memory ean therefore be interleaved in time.
26 All data transfers to and from memory by the 27 input/output ehannel are made by way of the system data map.
28 The system data map adds additional bits in the eourse of 29 translating the logieal addresses to physical addresses.
This permits a larger number of words of physieal memory z~

1 to be accessed by using a shorter logical address to access 2 a larger physical space than the word width itself would 3 normally allow.
4 The present invention also provides a syndrome decoding method for detecting and correcting errors in 6 semiconductor memory modul~s.
7 The storage area of the semiconductor memory 8 module comprises words of 22 bits. Each word has a 16 bit g data field and a six bit check field.
Each memory module includes an error de~ector for 11 simultaneously correcting all single bit and detecting all ~2 double bit errors and detecting many of the errors of 3 bits 13 or more anywhere in the 22 bit word. The error correction 14 includes a check bit generator, a check bit comparator, and 15 a syndrome decoder.
16 The check bit generator provides a code in which 17 each check bit is a linear combination of eight data bits 18 and in which each data bit is a component of exactly three l9 check bits.
The check bit comparator provides six output 21 syndrome bits. The input of each of the output syndrome 22 bits is eight data bits and one check bit.
23 The syndrome decoder interprets the value of the 2~ six output syndrome bits and identiies the presence or 25 absence of errors and the type o~ errors, if any, in the 26 22 bit word.
27 A data bit complementer is also provided for 28 inverting a single data bit error detected by the syndrome 29 decoder and thus correcting the error.

~36~

The semiconductor memory system is there~ore tolerant of single bit failures and can be operated with sinyle bit failures until such time as it is convenient to repair the memory.
~ultiprocessor system apparatus and methods which lncorporate the structure and techniques described above and which are effective to function as described above constitute further, specific objects o~ this invention.
According to one aspect of the present invention, there is provided a multiprocessor system of the kind in which separate processor modules operate concurrently and cooperatively for system control and application processing, said multiprocessor system comprising, a plurality of separate processor modules, interprocessor bus means including an interprocessor bus connecting each processor module for direct processor module to processor module signaling and data trans~er and wherein the interprocessor bus means are separate and distinct from an input/output system and from a memory system bus, each processor moduie having a central processing unit, a memory for that module, and interprocessor control means for connecting the processor module to the interprocessor bus for signaling and data transfer with another processor module, bus controller means operatively associated with the interprocessor bus and each interprocessor control means for determin.ing the priority of data trans~ers between any two processor modules over the interprocessor bus and ~or controll-ing transmissions over the interprocessor bus, each processor module including input~output channel means for transferring data bet~een the processor module and ~ 22 -21~

one or more device controllers, a plurality of peripheral devices, a plurality o~ device controllers for controlling the transfer of data between the processor modules and the peripheral devices, a plurality of ports ln each device controller, a plurality of input/output buses connecting each device controller for access by a plurality o~ different processor modules and wherein the input~output channel means, input/output buses and device controllers are operati~ely connected to provide, directly or in combination with the interprocessor bus means and a processor module, a path for data transfer between each processor module and any device controller.
According to another aspect of the present invention, there is provided a method of operating a plurality of separate processor modules concurrently and cooperatively for system control and application processing in a multiprocessor system of the kind in which each processor module includes a central processing unit and a memory ~or that module, said method comprising connecting each processor module in the system with each other processor module in the system for direc~ processor module to processor module si.gnalling and data kransfer by an interprocessor bus which is separake and distinct from an input/output system and from a memory system bus, connecting each processing module -to the interprocessor bus for signalling and data transfer with another module by an interprocessor control in the processor module, determining the priority of data transfers between any - 22a -~13167~28 two processor modules o~er the interprocessor bus and controll-ing transmissions over the interprocessor bus by a bus controller operatively associated with the interprocessor bus and with each interprocessor control of each processor module, trans~erring data ~etween a processor module and one or more device controllers connected to the processor module by an input/output channel in each processor module, connecting a plurality of peripheral devices to each device controller and controlling the trans~er of data between the processor modules and the peripheral devices by the device controllers, connecting each device controller for access by a plurality of different processor modules by a plurality of ports in each device controller and a plurality of input/output busses connected to the respective plurality o~ ports in each device controller, operatively connecting the input/output channels, input/output busses and device controllers to provide, directly or in combination with the interprocessor bus and a processor module, a path for data transfer between each processor module and any device controller so that a process running in any processor module can access any input~output resource in the system, whether or not the input/output resource is controlled by or attached to the processor module in which the process is running, by the process making re~uests comprising data transmissions over the interprocessor bus to a coop~rating process in a proce~sor module which controls that input/output resource.
The invention wlll now be described in greater detail with re~erence to the accompanying drawings, in which:

- 22h -Figure 1 is an isometric, block cliayram view of a multiprocessor system cons-tructed in accordance with one embodiment of the present invention. Figure 1 shows several processor modules 33 connected b~ two interprocessor buses 35 (an X bus and a Y bus) with each bus controlled by a bus controller 37. Figure 1 also shows several dual-port device controllers 41 with each device controller connected to the input/output (I/O) buses 39 of two processor modules;
Figure 2 is a block diagram view showing details of the connectlons of the X bus controller and the Y bus controller to the individual processor modules. Figure 2 shows, in diagram~atic form, the connections between each bus controller and the interprocessor control 55 of an individual processor module;
Figure 3 is a detailed diagrammatic view of the logic of one of the bus controllers 37 shown in Figure 2;
Figure 4 is a detailed diagrammatic view of the logic for the shared output buffer and control 67 in the interprocessor control 55 of a processor module as illus-trated in Figure 2;
Figure 5 is a view like Figure 4 but showing the logic for an inqueue buffer and control 65 of the inter-processor control 55 for a proce~sor modul~;

67~l 1 Fig. 6 is a state diagram of the logic 81 for 2 a bus controller 37 and illustratcs how the logic res-3 ponds to the protocol lines going into the bus controller 4 and generates the protocol lines going out of the bus 5 controller to the processor modules;

7 Fig. 7 is a state diagram like Fig. 6 but 8 showing the Iogic 73 and 75 for the shared outqueue 9 buffer and control 67 of Fig. 4 11 Fig. 8 is a state diagram like Figs. 6 and 7 12 but showing the logic 93 and 101 for the inqueue buffer 13 and control 65 of Fig. 5;

Fig. 9 is a diagrammatic view showing the time 16 sequence for the transmission of a given packet between 17 a sender processor module and a receiver processor 18 module;

Fig. 10 is a logic diagram of the bus empty 21 state logic section 75 and the processor fill state 22 logic section 73 of the outqueue buffer and the control --23 67 shown in Fig. 4;

Fig. 11 is a listing of logic equations for 26 the logic diagr~n shown in Fig, lO

28 Fig. 12 is a block dia~ram of the input/output 29 (I/0) system of the multiprocessor system shown in 30 Fig- l;

7~:~

1 Fig. 13 is a blocX diagram of the input/output 2 (I/O) channel 109 of a processor module. Fig~ 13 shows 3 the major components of the I/O channel and the data 4 path relating those component parts;

6 Fig. 14 is a detailed vie~l showing the 7 individual lines in the I/O bus 39 of Fig. l;
~ Fig. 15 is an I/O channel protocol diagram 10 showing the state changes of the T bus 153 for an execute 11 input/output (EIO) caused by the microprogram 115 in the 12 CPU 105. The sequence illustrated is initiated by the 13 CPU lOS and is transmitted through the I/O channel 109 14 of the processor modula 33 and on the T bus 153 to 15 device controller 41 as shown in Fig. l;

17 Fig. 16 is an I/O channel protocol diagram 18 showing ,the state changes of the T bus 153 for a reconnect 19 and data transfer sequence initiated by the I/O channel 20 microprogram 121 in response to a request signal from a 21 device controller 41;

23 Fig. 17 is an I/O channel protocol diagram 24 showing the state changes o~ the T bus 153 for an 25 interrogate I/O (IIO) instruction or an interrogate high 26 priority I/O (~IIIO) instruction initiated by the CPU
27 microprogram 115. The sequence illustrated is trans-28mitted over the T bus 153 to a de~Jice controller 41;

.

~L~L3&7~

1 Fig. 18 is a table identifying the functions 2 referred to by the mnemonics in Figs. 15 through 17;
4 Fig. 19 is a block diagram showing the 5 general structure of the ports 43 and a clevice controller 6 41 as illustrated in Fig. l;
8 Fig. 20 is a block diagram of a port 43 shown 9 in Fig. 19. This Fig. 20 shows primarily the data paths 10 within a port 43;

-12 Fig. 21 is a ~lock diagram showing the data 13 path details of the interface common logic 181 of the 14 device controller 41 shown in Fig. 19;
16 Fig. 22 is a block diagram showing the component 17 parts of a data buffer 189 in the control part of a 18 device controller 41 as illustrated in Fig. 19;

l 20 Fig. 23 is a graph illustrating the operation `~ 21 of the data buffer 189 illustrated in Figs. 22 and Fig. 19;

j 23 Fig. 24 is a timing diagram illustrating the 24 relationship of SERVICE OUT (SVO) ~rom the channel 109 . ~
25 to the loading of data into the poxt data register 213 .~ ..
` 26 ~Fig. 21) and illustrates how the parity check is started :~ 27 before data is loaded into the register and is continued 28 until after the data has been fully loaded into the register;

Fig. 2S is a schematic view showing details 31 of the power on circuit (PON) sho~m in Figs. 19 and 21;

., , ~3~7;~:~

1 ~ig. 26 is a logic dia~Jram o~ the buffer 2 control logic 243 of the data buffer 189 (shown in 3 Fig. 22) of a de~ice controller 41. Fig. 26 shows 4 how the buffer control logic 243 controls the hand-- 5 shakes on he data bus and controls the input and output 6 pointerS;

8 Fig~ 27 is a listing of the logic equations 9 for the select register 173 shown in Fig. 20. These 10 logic equations are implemented by the port control 11 logic 191 shown in Fig. 20;

13 Fig. 28 is a timing diagram showing the 14 operation of the t~Jo line handshake between the I/O
1~ channel 109 and the ports 43;

17 FigO 29 is a logic diagram showing the logic 18 for the general case of the handshake shown in Fig. 28.
19 The logic shown in Fig. 29 is part of the T bus machine 20 143 of the input/output channel 109 shown in Fig. 13;

22 Fig. 30 is a block diagram of a power - 23 distribution system. Fig. 30 shows how a plurality of 24 independent and separate power supplies 303 are 25 distributed and associated with the dual port device 26 controllers 41 ~or insuring that each device controller ~7 has both a primary and an alternate power supply;

2~ Fig. 31 is an enlarged, detailed view of 30 the switching arrangement for switching between a 1~3~7~

1 primary power suppl~ and an alternate supply for a 2 device contxoller. The switching structure shown in 3 Fig. 31 permits both auto~atic switching in the event of a failure of the primary po~7er sup~ly and manual S switching ln three different modes--off, auto and 6 alternate;

8 Fig. 32 is a block diagram showing details g of one of the separate and independent power supplies 10 303 illustrated in Fig. 30;

12 Fig. 33 is a block diagram ~iew showing 13 details of the vertical buses and the horizontal buses 14 for supplying power from the separate power supplies `15 303 shown in Fig. 30 to the individual device controllers 16 41. The particular bus arrangement shown in Fig. 33 17 permits easy selection of any two of the indi~idual 18 power supplies as the primary and the alternate power 19 supply for a particular device controller 21 Fig. 34 is a block diagram of the memory 22 system and shows details of the memory 107 of a processor 23 module 33 shown in Fig. 1;
~4 Fig. 35 is a block diagram showing details 26 of the map section 407 of the memory 107 shown in 27 Fig- 34;

29 Fig. 36 is a block diagram showing the 30 organization of logical memory into ~our logical address 2~

.. , ~L3~7Z8 1 areas and four separate map sections correspondir,g to 2 the four logical address areas. Fig. 36 also shows 3 details of the bits and fields ln a single map entry 4 -of a map section;

6 Fig. 37 is a ~lock diagram showing details of one of the memory modules 403 illustrated in Fig. 34.
8 The memory module 403 shown in Fig. 37 is a semiconductor g memory module;

~1 ~ Fig. 38 is a diagram of a check bit generator 12 used in the semiconductor memory module 403 shown in 13 Fig. 37. Fig. 38 also lists logic equations for two of 14 the eight bit parity trees used in the check bit register;

16 Fig. 39 is a diagram o a check bit comparator used in the semiconductor memory module 403 shown in 8 Fig. 37. Fig. 39 includes the logic equation for nine 19 bit parity tree for syndrome bit zero;

21 Fig. 40 is a diagram of a syndrome decoder ~2 used in the semiconductor memory module 403 shown in 23 Fig. 37. Fig. 37 also lists the logic equations for 2~ the operation of the logic section 511 o the syndrome 25 decoder;

27 Fig~ 41 is a logic diagram of a bit comp~.ementer 28 used in the semiconductor memory module 403 shown in 29 Fig. 37; and , _ .
. .

~L3~

1 Fig. 42 shows the various states of a t~o 2 processor system running an application program which is 3 required to be running continuously. The diagrams 4 illustrate the two processors successively Eailing and 5 being repalred and the application program chanying its 6 mode of operation accordingly.

~1 72~

THE MULTIPROCESSOF< SYSTEM-Eigure 1 is an isometric diagrammakic view of a part of a multiprocessor system constructed in accordance with one embodiment of the presen-t invention. In Figure 1 the multiprocessor system is indicated generally by the reference numeral 31.
The multiprocessor system 31 includes individual processor modules 33. Each processor module 33 comprises a central processing unit 105, a memory 107, an input/out-put channel 109 and an interprocessor control 55.
The individual processor modules are inter-connected by interprocessor buses 35 for interprocessor communications.
^ In a specific embodiment of the multiprocessor system 31, up to sixteen processor modules 33 are inter-connected by two interprocessor buses 35 (indicated as the X bus and the Y bus in Figure 1).
Each interprocessor bus has a bus controller 37 associated with that bus.

~IL1367Z~

1 The bus controllers 37, interprocessor buses 35 2 and interprocessor controls 55 (Fig. 1), together with 3 associated microprocessors 113, microprograms 115 and bus 4 receive tables lS0 (Fig. 2) provide an interprocessor bus system. The con~truction and opera~ion of this interprocessor 6 bus system are illustrated in Figs. 2 - 11 and 42 and are 7 described in more detail below under the subtitle The 8 Interprocessor Bus System.

The multiprocessor system 31 has an input/output 11 `(I/01 system for transferring data between the processor ~2 modules 33 and peripheral devices, such as the discs 45, 13 terminals a7, magnetic tape drives 49, card readers 51, and 14 line printers 53 shown in Fig. 1 16 The I/0 system includes one I/0 bus 3g associated 17 with each I/0 channel 109 o~ a processor module and one or 18 more multi-port device controllers 41 may be connected to 19 each I/0 bus 39, 21 In the specific embodiment illustrated, each device 22 controller 41 has two ports 43 for connection to two different 23 processor modules 33 so that each device controller is 24 connected or access by two processor modules.

26 The I/0 system includes a microprocessor 119 27 and a microprogram 121 in the I/0 channel 109 (See Fig. 12.) 28 which are dedicated to input/output transfers.

2g

32 :, ' ~' r 113~7Zf~

1 As also dia~rammatic~ illustrated in Fig. 12, 2 the microprocessor 113 and microprogram 115 of the c~ntral 3 processing unit 105 and an in~ut/output control table 140 in 4 the main memory 107 of each processor module 33 axe operatively associated with the I/O channel 109.

7 The construction and operation of these and other 8 components of the I/O system are illustrated in Figs. 12 - 2g g and are described in detail below under ~he subtitle The Input/Output System and Dual Port Device Controller.

11 , _,, 12 The multiprocessor system includes a power distribution 3 system 301 which distributes power from separate power supplies 14 to the processor modules 33 and to the device controllers 41 in a way that perrnits on-line maintenance and also provides 16 redundancy of power on each device controller.

18 As illustrated in Fig. 30, the power distribution 19 system includes separate and independent power supplies 303.

21 A separate power supply 303 is provided for each ~2 processor module 33, and a bus 305 supplies the power rom 23 the power supply 303 to the central processing unit 105 24 and memory 105 of a related processor module 33.

33 ~l36~

As also illustrated in Figure 30, each device controller 41 is connected for supply of power ~rom two separate power supplies 303 throuyh an automatic switch 311. If one power supply 303 for a particular device con-troller 41 fails, that device controller i9 supplied with power from the other power supply 303; and the changeover is accomplished smoothly and without any interruption or pulsation in the power supplied to the device controller.
The power distribution system coacts with the dual port system of the device controller to provide un-interrupted operation and access to the peripheral devices in the event of a failure of either a single port 43 or a single power supply 303.
The multiprocessor system includes a power on (PON) circuit 182-(the details of which are shown in Figure 25) in several components of the sys-tem to establish that the power to that particular component is within certain acceptable limits.
For example, the PON circuit 182 is located in each CPU 105, in each device controller 41, and in each bus controller 37.

- 34 -~L3~2~

1 ~he purpose of the PON circuit is to present 2 a signal establishing the level of power applied to that 3 particular component; and if the power is not within 4 certain predetermined acceptable limits, then the siynal 5 output is ~sed to directly disa~le the appropriate bus 6 signal of the component in which the PON is located.

8 The power-on circuit functions in four states --g power off; power going from off to on; power on; and 10 power going from on to off.

12 ~he power-on circuit initializes all of the logic 13 states of the system as the power is brought up; and in 14 the present invention, the power-on circuit provides an 15 additional and very important function of providing for 16 a fail-safe system with on line maintenance. To do this, 17 the power-on circuit in the present invention is used in 18 a unique way to contxol the interface circuits which dri~e 19 all of the intercommunication buses in the system.

21 The construction and operation of the power 22 distribution syskem are illustrated in Figs. 30-33 and 23 are described in detail below under the subtitle Power 24 Distribution System.

2.5 26 The multiprocessor system includes a memory system 27 in which the physical memory is divided into four logical ~8 address areas -- user data, system data, user code and 29 system code (See Fig. 36.)~

, 1~l;3 &~8 1 The memor~ system includes a m~p 407 and control 2 logic 401 (See Fig. 34.) for translating all logical addresses 3 o physical addresses and for indicating pages absent from 4 primary storage bit present in secondary storage as réquired 5 to implement a virtual memor~ system in which the physical 6 page addresses are invisible to users.

8 The memory system incorporates a dual port access g to the memory by the central processing unit 105 and the 0 I/O channel 109. The I/O channel 109 can therefore access 11 the memory 107 directly (without having to go through the 12 central processing unit 105) for data transfers to and from 13 a device controller 41.

lS The construction and operation of the memory 16 system are illustrated in Figs. 34-41 and are described in 17 detail below under the subtitle Memory System.

19 An error detection system is incorporated in 20 the memory system for correcting all single bit and detecting 21 all double bit errors when semiconductor memory is used in 22 the memory system. This error detection system utilizes a 23 16 bit data field and a 6 bit check field as shown in Fig. 37 24 and includes a data~bit complementer 487 as also shown in ~` 25 Fig. 37 for correcting single bit errors.

,..
~ `

~3~7Z~

1 Figs. 37 through 41 and the related disclosure 2 illustrate and describe details of the error detection 3 system.

Before going into the detailed description of 6 the systems and components noted generally above, it should 7 be noted that certain terminology ~lill have the following 8 meanings as used in this application.

The term "soft~are" will refer to an operating 11 system or a user pro~ram instructions; the term "firmware"
12 will refer to a microprogram in read only memory; and 13 the term "hardware" will ref~r to actual electronic logic 14 and data storage.

16 The operating system is a master control program 17 executing in each processor module which has primary control 18 Of the allocation of all system resources accessible to 19 that processor module. The operating system provides a 20 scheduling function and determines what process has use of 21 that processor module. The operating system also allocates ~2 the use of primary memory (memory management), and it 23 operates the file system for secondary memory management.
24 The operating system also manages the message system.
2~ This provides a facility for information transfer over 26 the interprocessor bus.

1 The operating system arran~ement parallels 2 the modular arrangement of the multiprocessor syst~m 3 components described above, in that there are no "global"
4 components.
6 At the lc~est level of the software system, 7 two fundamental entities are implemented--processes and 8 messages.

A process is the fundamental entity of control 11 within a system.

13 Each process consists of a private data space 14 and register values, and a possibly shared code set. A
15 process may also access a common data space.
i~ .
17 A number of processes coexist in a processor 18 module 33.
19 .
The processes may be user written programs, or 21 the processes may have dedicated functions, such as, for 22 example, control of an I/O device or the creation and 23 deletion of other processes.

A process may request services from another process, and this other process may be located in the same 27processor module 33 as a process making the request, or 28the other process may be located in some other processor 29module 33.

3a i7;Z8 1 ~he processes ~ork in an as~nchronous manner, 2 and the processes therefor~ nced a method of communicatiOn 3 that will allow a request for services to be queued with 4 out "races" (a condition in which the outcome depends 5 upon the sequence of which process started first) -thus 6 the need for "messages" (an orderly system of interprocessor .7 module communication described in more detail below).

9 Also, all interprocessor module communication should appear the same to the processes, regardless of 11 whether the processes are in the same or in differen~-12 processor modules.

14 As will become more clear from the description 15 to follow, t.he software structure parallels the hardware;
16 and different processes can be considered equivalent to 17 certain components of the hardware in arrangement and 18 function.

For example, just as the I/O channel lO9 21 communicates over the I/0 bus 39 to the device controller 22 41, a user process can make a request (using the message . 23 system) to the process associated with that device controller : 24 41; and then the device process return9 status back 25 similar to the way the device controller 41 returns 26 information back to the I/O channel 109 over the I/0 bus 39.

28 The other fundamental entity of the software 29 system, the message, consists of a request for service as 30 well as any required data. When the request i5 completed, 3g ~L3~7~

any required values wlll be returned to the requesting 2 process.
4 When a message is to be sent between processes 5 in two dif~erent processor modules 33, t~e interproce~sor 6 buses 35 are used. However, as noted above, all communication 7 between processes appears the same to the processes, 8 regardless of whether they are in the same or in different 9 processor modules 33.

11 This software organization provides a number 12 of benefits.

14 This method of structuring the software also 15 pro~ides for significantly more reliable software. By 16 being able to compartmentalize the software structure, 17 smaller module sizes can be obtained, and the interfaces 18 between modules are well defined.
. 19 i ` 20 The system is also more maintainable because 21 of the compartmentalization of function.
2~ -23 The well defined modules and the well defined 24 interfaces in the software system also provide advantages 25 in beiny able to make it easily expandible--as in the 26 case of adding additional processor modules 33 or device 27 controllers 41 to the multiprocessor sy~tem.

29 Furthermore, there is a bene~it to the user 30 Of the multiprocessor system and software system in that ~0 ~3~

1 the user, writing his pro~ram, need not be a~are of either 2 the actual machine configuration or the ph~sical location 3 of other processes.

, Just as the hard~are provides multiple function-~ ally equivalent modules ~ith redundant interco~nects, so 7 does the software.

9 For example, messages going between processes 10 in different processor modules 33 may use either inter-11 processor bus 35. Also, device controllers 41 may be 12 operated by processes in either of the processor modules 13 33 connected to the device controller 41.

The multiprocessor hardware system and software l~ system described above enable the user to develop a 17 fault tolerant application system by virtue of its 18 replicated modules with redundant interconnects.

~L367~

1 THE INTERPROCESSOR BUS SYSTEM:

3 As pointed out above, ~he individual processor 4 modules 33 are interconnected by two interprocessor buses 5 35 ~an X bus and a Y bus) with each bus controlled by a 6 related bus controller 37. Each interprocessor bus 35, in 7 combination with its bus controller 37 and a related inter-8 processor control 55 in each processor module 33 r provides g a multi-module communication path fxom any one processor 10 module to any other processor module in the system. The 11 use of two buses assures that two independent paths exist 12 between all processor modules in the system. Thereore, a 13 failure in one path (one bus) does not prevent communication 14 between the processor modules.
16 The bus controller 37 for each interprocessor bus 17 35 is a controller which is, in a preferred form of the 18 invention, separate and distinct from the processor modules 33.

Each interprocessor bus 35 is a synchronous 21 bus with the time synchronization provided by a bus clock 22 generator in the bus controllers 37. The interprocessor 23 control portions 55 of all of the modules associated with 24 the bus make state changes in synchronism with that bus 25 clock during transfers over the bus.

27 As will be described in more detail below, the 28 CPU 105 operates on a dif~erent clock from the inter-~9 processor bus clock. During the filling of an outqueue 30 or the emptying of an inqueue in the interprocessor control 4~

~L3il3~;7~

1 55 by the CPU, the operation takes place at the CPU clock 2 rate. However, transmission of packets over ~he inter-3 processor bus always takes place at the bus cloc~ rate.

~t is an important feature of the present 6 invention that the information transmitted over the inter-7 processor bus is transferr~d at high~transmission rates 8 without any required correspondence to the clock rates of 9 the various CPUs 105. The information transfer rate over 10 the interpxocessor bus is also substantially faster than 11 would be permitted by direct memory accesses into and out 12 of the memory sections 107 at memory speed. ~his ensures 13 that there is adequate bus bandwidth even when a large number 14 of processor modules is connected in a multiprocessor system.

16 A benefit of using separate clocks for each CPU
17 105 is that a master system clock is not required, and 18 this eliminates a potential source of single component 19 failure which could stop the entire system.

~ 21 The interprocessor control 55 incorporates logic i~ 22 interlocks which make it possible to operate the inter-. .
- 23 processor buses 35 at one clock rate and each CPU 105 24 at its own independent clock rate without lo.~ of data.
.
~ 25 -~ 26 ~he information transmitted over the bus is 27 transmitted in multiword packets. In a preferred form ` 28 Of the present invention each packet is a sixteen word ; .
29 packet in which fifteen o the words are data words and 30 one word is a check word.
.
.
~3 ~ .
~, ..
~ . -~136~7Z8 1 The control logic within the ~us controller2 37 and the interprocessor controls 55 of the individual 3 modules 33 follows a detailed protocol. The protocol 4 provides for establishing a sender-receiver pair and a 5 time frame~ for the data packet transfer. At the end of 6 the time frame for the transmission of the data packet, 7 the bus controller 37 is released for another such sequence.
8 The specific manner in which these functions are carried 9 out will become more apparent after a description of the 10 structural features of Figs. 3-9 below.

11 .. ..
12 X bus 35 is identical in structure to the 13 Y bus 35, so the structure of only one bus will be 14 dPscribed in detail.
1~ .
16 As illustrated in Fig. 2, each bus 35 comprises 17 sixteen individual bus data lines 57, five individual 18 bus pro~ocol lines 59, and one clock line 61, and one 19 select line 63 for each processor module 33.

2~ As also illustrated in Fig. 2, the inter-22 processor control 55 of each processor module 33 includes 23 two inqueue sections 65 (shown as an X inqueue section 24 and a Y inqueue section in Fig. 2) and a shared outqueue 25 section 67.

27 With the specific reerence to Fig. 4, the 28shared outqueue section 67 includes an outqueue buffer 69 29which performs a storage function. In a prefe~red form 30the buffer 69 has sixteen words of sixteen bits each. The ~3~

1 buffer 69 is loaàed by the CPU and holds the data until the 2 packet transmission time, at which time the d~ta is gated 3 out to the bus, as ~7ill be described in more detail below.
The outqueue section 67 also includes a receive 6 register 71, which in a preferred form of the invention 7 is a four bit register. This register is loaded by the 8 CPU with the number of the processor module to which the 9 data will be sent.
11 The control part of the outqueue section 67 12 includes a processor fill state logic section 73 which 13 operates in synchronism with the CPU clock, a bus empty 14-state logic section 75 which operates in synchronism with 15 the X or Y bus clock, and an outqueue counter 77. During 16 filling o~ the outqueue buffer 69 by the CPU, the out-17 queue counter 77 scans the buffer 69 to direct the data 18 input into each of the sixteen words of the buffer; and, 19 as the sixteenth word is stored into the outqueue buffer 20 69, the OUtqUQUe counter 77 terminates the fill state.

22 The outqueue section 67 also includes an out-23 queue pointer 79 which connects the entire outqueue 24 section to either the X bus or the Y bus 35. The outqueue 25 pointer 79 allows the logic sections 73 and 75 and the 26 bufer 69 to be shared by the X and Y interprocessor buses 35.

28 As illustrated in Fig. 3, the bus controller 29 37 comprises a bus control state logic section 81, a 30 sender counter 83, a processor select logic section 85, ~5 ~L~36~Z8 1 a receive register 87, a packet counter 89 and a bus 2 clock generator 91.

4 With xeference to Fig. 5, each inqueue section 65 comprises a bus fill state logic section 93 which 6 operates in synchr~nism with the bus clock, a sender 7 register 95, an inqueue buffer 97, an inqueue counter 99, 8 and a processor empty state logic section 101 which 9 operates in synchronism with the CPU clock.

11 Fig. 6 is a state diagram of the bus control 12 logic 81 of the bus controller 37.

14 Figc 7 is a state diagram o the logic sections 15 73 and 75 of the outqueue section 67.
1~
17 Fig. 8 is a state diagram of the logic sections 18 93 and 101 of the inqueue sections 65.
'19 With reference to Fig. 7, the processor fill state 21 logic section 73 has basically four states--EMPTY, FILL, FULL
22 and WAI~--as indicated by the respective legends. The bus 23 empty state logic section 75 has basically four states--~4 IDLE, SYNC, SEND and DONE--as illustrated by the legends.
. :

26 Continuing with a description of the notation in 27 Fig. 7, the solid lines with arrows lndicate transitions 2g from the present state to the next state. Dashed arrows 29 ending on the solid arrows indicate conditions which must 30 be satisfied for the indicated transition to take place.
.

` 46 72 !3 1 The synchroniz~tion of state machines running 2 off relatively asynchronous clocks require a careful 3 construction of an interlock system. These important 4 interlocks are noted by the dashed arrows in the state 5 diagrams. -These interlocks perform a synchronization 6 of two relatively asynchronous state machines. The 7 dashed arrows in Fig. 7 and Fig. 8 running ~etween the 8 state machines thus indicate signals which synchronize g (qualify) the indicated transistions of the state machines.

11 ` , ' ~2 ~i 24 `~ 26 , ~ 27 `~ 28 ' .

I

113~7~

1 With reerence to the FILL state fox the 2 logic section 73, 't should be noted that the store 3 outqueue condition will not cause an exit from the 4 FILL state until the outqueue counter 77 has advanced S to count 15 (on a count which starts with zero) 6 at which time the ~ILL state will advance to the FULL
state.
9 Similarly, it should be noted that the SEND
10 state of the logic section 75 will not terminate on the 11 select and send co~mand condition until the outqueue 12 counter 77 reaches count 15, at which time the SEND
13 state advances to the DONE state.

The asterisk in the notation of Fig. 7 16 indicates an increment of the outqueue counter 77.

18 Fig. 6 shows the state diagram for logic 81 19 of the bus controller and illustrates that the logic 20 has basically four states- IDLE, POLL, RECEIVE and SEND.
~21 22 - The notation in Fig. 6 is the same as that 23 described above for Fig. 7~ A solid arrow }ine indicates 24 a state transition ~rom one state to another and a 25 dotted arrow line to that soli.d arrow line indicates a 26 condition which must occur to allow the indicated 27 (solid line arrow) transition to occur. An asterisk 2~ on a state transition in this case indicates that 29 simultaneously with the indicated transition the sender 30 counter 83 lS incremented by one.

6~

.. ~.-`' ' I

7~

1 The dashed arro~ output lin~s in Fic3. 6 2 indicate protocol commands issued from the bus 3 controller to the interproc~ssor bus.

~n both Fiy. 7 and Fig. 6 a dashed arrow 6 leaving a state indicates a logic output from that 7 state such as a logic output signal to a protocol line 8 (in the case of the bus empty state logic 75) or to a 9 status line of the processor module (in the case of the 10 processor fill state logic 73).

11 ' 12 Fig. 8 shows the state diagrams for the bus 13 fill state logic section g3 and the processor empty state 14 logic section 101.

16 The state diagram for the logic section 93 17 includes four states--SYNC, ACKNOWLEDGE, RECEIVE and FULL.

19 The state diagram for the logic section lOl 20 includes four states--RESET, READY, INTERRUPT and DUMP.

22 The notation ~solid line arrows and dashed line ~3 arrows) is the same as described above for Fig. 7 and Fig. 6.

The asterisk in Flg. 8 indicates an increment 26 in the inqueue counter g9.

28 Fig. 9 is a timing diagram showing the time 29 sequence in which the state changes given in Figs. 6, 7 30 and 8 occur~

~,9 7~9 1 The sequence shown in Fig. 9 accomplishes 2 the transmission of a packet from one proc~ssor modulc 3 to another processor module at the bus clock rate 4 (assuming that the intended receiver module is ready to 5 receive the packet).

7 Fig. 9 shows the time sequences for a success-8 ful packet transfer with individual signal repre~.entations 9 listed from top-to-bottom in Fig. 9 and with time periods 10 of one bus clock each shown from left-to-right in the ll order of increasing time in Fig. 9.

13 The top line in Fig. 9 indicates the state of 14 the bus controller, and each division mark represents a 15 clock period or cycle of the bus clock generator 91 shown 16 in Fig. 3. Each time division of the top line carries 17 down vertically through the various signal representations 18 listed by the legends at the left side of the figure.
lg ~ .
Taking the signals in the sequence presented 21 from top-to-bottom in Fig. 9, the first signal (below 22 the bus controller state line) is the SEND REQUEST signal 23 (one of the protocol group indicated by the reference 24 numeral 59 in Fig. 3) and specifically is the signal 25 which may be asserted by the outqueue control logic 26 section 67 of any processor module 33. The signal is 27 transmitted to the bus control state logic section 81 , . . .
2 2~ of the bus controller 37 (see Fig. 3), ~ 29 .~ 30 `
~`, .
~0 1 The next signal shown in Fig. 9 (the SELECT
2 signal) represents a signal which originates rom the 3 processor select logic section 85 of the bus controller 4 37 and which is transmitted on only one at a ~ime of the 5 select lines 63 to a related processor module 33.

7 The next signal represented in Fig. 9, the 8 SEND ACKNOWLEDGE signal, may be asserted only by a g particular processor 33 when that processor is selected 10 and when its bus empty state logic section 75 is in 11 the SEND state (as illustrated in the third state of 12 Fig. 7). This SEND ACKNOWLEDGE signal is used by the 13 bus controller 37 to establish the identity of a processor 14 module 33 wishing to send a packet.
16 ` The next signal, the RECEI~E COMMAND signal, repre-17 sents a signal from the bus controller 37 transmitted on one 18 of the protocol lines 59. This signal does two things.

First of all, this signal in combination with 21 receiver SELECT interrogates the receiver processor module 22 33 to find out whether this receiver module is ready to 23 receive (as indicated by the ACKNOWLEDGE state in Fig. 8).

Secondly, this signal ha5 a secondary function 26 Of disabling the ~U5 empty state logic section 75 of the 27 receiving module so that the receiving module cannot gate 28 an intended receiver number to the data bus should the 29 out~ueue section of khe intended receiver module 33 ~lso 30 have a data packet of its own ready to send.

~ IL3Ei7~

1 In this regard, during the ~ime that the 2 sender processor is asserting the SEND ACKNO~1LEDGE
3 signal it is also gating the receiver number to the 4 bus for use by the bus controller 37~ The bus 35 itself is, of course, a non-directional ~us so that the 6 information can be gated to the data bus 57 by any ~odule 7 for use by either the bus controller 37 for a control 8 function or for use by another processor for an information g transfer function. It should be noted that a module 33 0 may gate data to the bus only when its SELECT line is 11 asserted and the RECEIVE COMMAND signal is not asserted.

13 During the time that the RECEIVE COMMAND signal 14 is asserted the bus controller 37 is ga~ing the sender number to the data bus 57 for capture by the selected 16 receiver processor module.

18 The next signal line (the RECEIVE ACKNOWLEDGE
19 line in Fi~. 9) represents a signal which is transmitted 20 from the selected receiving module's bus fill state logic 21 section 93 to the bus control state logic section 81 of 22 the bus controller 37 (over one of the protocol lines 59) 23 to indicate that the 5elected receiver module is in the 24 Ac~NowLEDGE state (as indicated b~ the legend in Fi~. 8) 25 and thus ready to receive the packet which the sender 26 module has ready to transmit.

28 If the RECEXVE ACKNOWLÉDGE signal is not 29 asserted by the receiver module, the sender SELECT, 30 the SEND CO~MAND and the time frame transmission of 31 the data packet itself will not occur.

;7;~8 1 If the RECEIVE ACK~IOWLEDGE signal is asserted, 2 then the sequence indicated by the SEND CO~MD line 3 will occur.

The SEND CO~ AND line represents a signal 6 which originates from the bus control state logic 7 section 81 of the bus controller 37 and which is trans-8 mitted to the bus empty state logic section 75 of the ~ sender processor module 33 over one of ~he protocol lines 1~ 59..

11 ' 1~ In combination with a SELECT o the sender 13 processor module the SEND COMMA~D signal enables the 14 sender processor module to send a packet to the 1~ receiver module during the sixteen clock cycles 16 bracketed by the SEND COM~ND signal.

18 The final line (the data/16 line) represents 19 the information present on the data lines 57 during the 2Qabove-described sequence.

22 The data is gated to the bus by the selected sender 23processor module and is transmitted to the receiver 24processor module into the inqueue buffer 97 ~see Fi~. 5) 5during this sixteen clock cycle time frame. This assumes 26that the RECEIVE ACKNOWLEDGE signal wa~ recei~ed by the 27bus controller in response to the RECEI~E COMMAND signal~

. ~ , , .
' ' 29 ~- 30 ~: 53 ~ .
, - , . .

&~

1 If the REOEIVE AC~NOWLEDG~ signal had not 2 been received by the bus controller, then the SEND
3 CO~ND signal would not have been asserted and the 4 bus controller 37 would have resumed the POLL state 5 as shown in Fig. 6.
7 With reference to Figs. 2,. 7, 10 and 11, a 8 typical operation of the outqueue buffer and control 67 ~ of one processor module 33 will now be described.
11 As illustrated in Fig. 10, the processor fill 12 state logic section 73 includes two flip-flops A and B, 13 and the bus empty state logic section 75 includes two 14 flip-flops C and D.
16 Summarizing the state assignments as shown by 17 the AB and CD tables in Fig~ 10, the EMPTY state is 18 defined as A = 0, B = 0. The FILL state is defined as 9 A = 1, B = 0 . The FULL state is defined as A = l, B - l; .
20 and the WAIT state is defined as A = 0, B = 1.
~ ! 21 22 Similarly, the corresponding combinations of 23 the C and D state variablés are defined to be the IDLE, ~-~ 24 SYNC, SEND and DONE states res~ectively. State assign-~ 25 ments previously listed could also be gi~en in form o~
3 26 logic equations. Fox example, EMP'rY = A ~, and this ;, , ~ 27 notation is utilized in the Fig. 11 logic equation `~ 28 listings~
.~ .

. . 30 , 5~
,, , ~, ~'' '' .
, .

~3~7:~

In operation and with speci~ic reference to Figure 7, the initial state reached through power on initialization or manual reset i5 the EMPTY state shown in the top left part of Figure 7.
The EMPTY state of the processor fill state logic 73 provides a ready signal to the central processor unit (CPU) 105 to indicate the presence of that state, as indicated by the dashed arrow RDY shown as leaving the empty state in Figure 7.
The CPU firmware (microprogram) in response to that ready signal, when a transmission over the inter-processor bus is required, will provide a store receive signal (shown by the dashed arrow incoming to the diagram in Figure 7). This store receive signal qualifies (synchronizes) the transition which advances the EMPTY
state to the FILL state.
The CPU firmware, to transfer data into the outqueue buffer 69, will provide a store outqueue signal (the dashed arrow entering the diagram in Figure 7) for each word to be stored in the buffer 69.
Each occurrence of this store outqueue signal will advance the outqueue counter 77, commencing with a count of zero, until a count of 15 is reached.
On the sixteenth occurrence of the store out-queue signal a transition from the FILL to the ~'ULL state, 1 as illustrated by the solid line arrow in Flg. 7, is 2 allowed.
4 The FULL state of the processor FILL state logic 5 provides a synchronization condition to the bus empty state 6 logic denoted by th2 dashed arrow leaving the FULL state 7 of logic 73 and going down to the logic 75 in Fig. 7.

9 The processor fill state logic 73 will remain 10 in the FULL state until the bus empty state logic 7S
11 has subsequently reached the DONE state.

13 Now, referring specifically to the bus empty 14 state l~gic denoted by 75 in Fig. 7, the initial state, 15 IDLE, for ~he logic section 75 in Fig~ 7 is again pro-16 vided ~y power on initialization or manual reset.

18 ~The bus empty state logic 75 will remain in 19 the IDLE state until the transistion to the SYNC state is 20 allowed as shown by the dashed arrow from the FULL state 21 Of the processor fill 73.
. .
~i -22 . ~ .
23 The empty state logic 75 will proceed with no ~24 qualification required from the SYNC state to the SEMD
-25 state.

, 27 It is in the SEN~ state that the 5END REQUEST
28 signal to the bus and to the bus controller is asserted 29 (as indicated by the dashed arrow going down and leaving 30 the diagram 75 from the SEND state).

. 56 .- :
.~.- i , ~

In response to this SND REQUEST siynal, the 2 bus controller logic 81 (Fig~ 6) will poll processor 3 modules successively until the sender is identified - 4 (as discussed earlier with reference to Fig. 9).

h The bus controller will issue a RECEIVE COI~MAND
7 and SELECT to the intended receiver processor module; and 8 upon receipt,of the RECEIVE ACKNOWLEDGE signal will proceed 9 to the packet time,frame (also identified in Fig. 9).

11 During the packet time frame the bus controller 12 asserts SELECT of the sender processor module and also 13 asserts the SEND COMMA'ND signal to the sender processor module.

This SELECT signal and SEND COMMAND signal is 16 shown as entering the diagram and qualifying (synchronizing) 17 transitions leaving and entering the SEND state as noted 18 in Fig. 7.

Each bus clock while SELECT and SEND COL~AND
21 are asserted will advance the outqueue counter 77 commenc-22 ing with a count of zero.
- `23 24 On the sixteenth clock period of SELEC~ and SEND
25 COMMAND the transition terminating the SEND state and ad-2~ vancing to the DONE state is qualiied tsynchronized as 27 shown by the dashed arrow allowing that transition).

' 29 When the empty state logic 75 has reached the 30 DONE state, a transition of the processor fill state logic , 57 ~L3~2~

73 Srom FULL to WAIT is qualified (as denoted by the dashed arrow leaving the done state).
Next, the WAIT state of the processor fill state logic 73 qualifies a transition of the bus empty state logic 75 from the DONE state to the IDLE state (as denoted by a dashed arrow leaving the WAIT state and qualif~ing the indicated transition).
Finally, the bus empty state logic 75, being in the IDLE state, qualifies the transition of the processor fill state logic 73 from the WAIT state to the EMPTY state (as denoted by the dashed arrow leaving the IDLE state).
At this point a packet has been loaded into the outqueue buffer 69 by the processor module and transmitted over the bus 35 to the receiver processor module, and the outqueue control processor fill state logic 73 and bus empty state logic 75 have returned to their initial states.
T~e above description relates to the transitions and qualifications indicated in Figure 7. The action of the logic sections 73 and 75 involved in the above des-cription of operation of Figure 7 will now be noted with reference to the logic diagram of Figure 10 and the logic equation listing of Figure 11.
~7ith reference to Flgure 10, as noted above, the flip-flops A and B are JK flip-flops and are edge triggered flip flops in that state chancJes occur only 72~

1 on clock transitions tas indicated by the small triangular 2 symbols and legends on the lefthar~d sides of the flip-flops 3 A and B in Fig. 10).

The primary significance of the logic 6 diagram in Fig. 10 is to lllustrate the transition from 7 one state to another in the state machines shown in 8 Fig. 7. Thus, to illustrate the transition from IDLE
9 to SYNC in the empty state logic 75, the operation 10 proceeds as follows.

1~
12 To implement a change from the IDLE state 13 to the SYNC state, the state variable C must be set.

The logic equation for the J input of state 16 variable C is as shown in Fig. 11 and is indicated by 17 the reference numer~l 103. In this equation the inter-18 lock ~shown by the dashed arrow from the full state of l9 the fill state logic 73 in Fig. 7 to the transition) 20 corresponds to the quantity ~A B) or (FULL) in the 21 equation indicated by the reference number 103. The D
22 or ~IDLE) in the equation indicated by reference nu~eral 23 103 in Fig. ll corresponds to the IDLE state shown by the 24 legend in Fig. 7. ~he J in the equation corresponds to 25 the J input of the C flip-flop in Fig. 10. And the ~C) 26 corresponds to the true output of the C flip-flop in 27 Fig. 10.
2~
29 Other state transitions of the Fig. 7 30 diagram will not be described in further detail with 7~

1 reference to Figs. 10 and 11 since it is believed 2 that these transitions as carried out by the logic 3 diagram in Fig. 10 and the logic equations in Fiy. 11 4 are clear from the above examples o the transitlon from IDLE state to SYNC state as described in detail above.

7 Figs. 10 and 11 show the logic diagram and 8 logic equations for the state diagram of the outqueue 9 buffer and control 67. Corresponding logic diagrams ~0 and logic equations have not been illustrated for the 11 inqueue huffer and control 65 or the bus controller 12 37 ~ecause such logic diagrams and equations are similar 13 to those shown in Fig. 10 and Fig. 11 and are easily 14 obtainable from the state diagrams shown in Figs. 6 and 8.

~6 Each processor module 33 (Fig. l) in the multi-17 processor system is connected to both interprocessor buses 18 35 (Fig: 1) and is capable of communicating with any pro-19 cessor module including itself o~er either bus. For each 20 block data transfer, one processor module is the source 2~ or sender and another is the destination or receiver.
~` 22 23 Transmission of data by a processor module ~- 2b~ over one of the interprocessor buses is initiated and ~ 25 accomplished under software control by means of the SEND
:
"~ 26 instruction.

`` 27 28 In the SEND instruction the microprogram 115 29 (Fig. 2) and the CPU microprocessor 113 (Fig. 2) interacts 30 with the shared outqueue section 67 of the interprocessor 1 .

1 control 55 to read a data hlock from memor~ lOl,to hreak 2 it up into packets, to calculate packet check sum ~70rds, 3 and to ~ransmit ~he block one packet at a time over a 4 bus to the receiving processor module. Parameters supplied to the SEND instruction specify the number of ~ words in the block, the starting address of the block, i 7 which bus to use, the destination processor, and a 8 maximum initial timeout value to wait for the outqueue 9 67 (Fig. 2) to become available.

11 The SEND instruction terminates only after the 12 entire block has been transmitted; thus sending a block 13 is a single event from the software viewpoint. However, 14 the SEND instruction is interrupta~le and resumable, so that responSe OL the operating system to other events is 16 not impaired by the length of the time required to 17 complete a SEND instruction.

19 Receiving of data by a processor module over the interprocessor buses is not done by means of a soft-21 ware instruction, since the arrival times and sources - 22 of data packets cannot be predicted. The receiving of z3 data i5 enabled but cannot be initiated by the receiver.
~" :
~; 24 ~` 25 ~he CPU microprocessor 113 takes time~ out from `~^ 26 software instruction processlng as required to execute . ~ .
` 27 the BUS RECEIVE microprogram 115. This microprogram 28 takes the received data packet from one o the inque~ue 29 sections 65 (Fig. 2) of the interprocessor control 55, stores the data into a memory buffer, a~d verifies correct 31 packet check sum.

., ~, . .
~ `

~3~8 1 Reassemhly of received packets into blocks - 2 is accomplished using the Bus Receive Table 150 (BRT) 3 in memory. The ERT contains 32 two-word entri~s, corre~-4 ponding to the two buses from each of the sixteen pro-cessor modules possible in one specif ic implementation 6 of the multiprocessor system. Each BRT entry corres-7 pondlng to a bus and a sender contains an address word 8 and a count word. The address word specifies into which 9 buffer in the System Data area incoming data from that sender is to be stored. The count word specifies hcw many 11 data words remain to complete the block transfer from 12 that sender.

14 As each data pac~et is received, the CPU micro-15 processor 113 suspends processing of software instructions, 16 and the bus receive microprogram 115 is activated. This 17 microprogram reads the address and count words from the 18 sender's BRT entry, stores the data packet into the 19 specified area, verifies correct packet check sum, and 20 restores adjusted values of the address and count words 2~ into the BRT entry. If the packet caused the count to 22 reach zero or if the packet contained incorrect check sum, 23 the bus receive microprogram sets a completion interrupt ` 24 flag to signal termination of the data block to the soft-25 ware. The CPU microprogram then resumes software 26 instruction processing at the point where it left o~f 27 with no disturbance except délay to the currently executing 28 program.

~1367~

1 It is an important feature that data blocks 2 from several senders can all be assembled concurrently 3 by a receivincJ processor module from data p~ckets received 4 in any sequence. This interleaved assembly of blocks from packets is carried on transparently to the soft-6 ware executing in ~he receiver processor. Only success~
7 ful block completions or erroneous transmissions cause the 8 software to be interrupted.

It is also important that a time-sharing or 11 time-slicing of the interprocessor bus hardware has been 12 achie~ed in two areas.

14 First, each interprocessor bus and associated bus controller allow packets to be transmitted between 16 any sender and receiver as required. The circular polling 17 by a bus controller to identify a requesting sender 18 ensures that all pxocessor modules have an equal opportunity 19 to send over that bus. Each bus provides a communication path which is shared in time in an unbiased way by all 21 processor modules.

.~ .
23 Secondly, each inqueue section 65 o~ the inter-24 processor control 55 of a processor module is shared in time by incoming packets from several senders. That is, 26 the inqueue logic and storage of a processor is not 27 dedicated to a single sender for the duration of a block 8 transfer. Instead, each packet received i3 correctly 29 directed into memory by the BRT entry corresponding to its sender and bus. Data blocks from several senders .

~; ~

~3~

1 are assembled correctly in the receiv~r's memory 2 independently of the order in which the senders make 3 use of the bus.

A processor module has two ~a~fs of controlling 6 its ability to rec~ive packets over the X bus or the Y bus.

8 First, there is a bit in the CPU's interrupt 9 MASK register corresponding to each interprocessor bus.

When the MASK bit is on, micro-interrupts for that bus 11 are allowed. Micro-interrupts (activation of the BUS

12 RECEIVE microprogram) occur when the Processor Empty 13 state logic 101 ~Fig. 5) of an inqueue section 65 reaches 14 the MICRO-INT state after a packet has been received into an inqueue buffer. If the MASK bit is off when a 16 packet is received, the micro-interrupt and subsequent 17 processing of the packet into memory will be deferred 18 until the MASK bit is set on by a software instruction.

Software operations such as changing a BRT
21 entry are performed with micro-interrupts disabled to 22 avoid unpredictable results. No packets are lost while 23 micro-interrupts are disabled. The firsk packet received 24 will be held in the inqueue buffer until the micro interrupt is enabled. Subsequent packet trans~ers while 26 the inqueue bufer is ~ull are rejected since the Bus ;
27 Fill state 93 logic will be in the FULL state and thus 28 unable to assert RECEIVE ACKNOWLEDGE in response to 29 SELECT.
3~

~. .

~L13~72~

A second means of controlliny its ~bili'c~ to 2 receive packets over the bus is the action taken by a 3 processor module after an X bus or Y bus receive 4 completion interrupt (activation of an operating system - 5 interrupt handler).

7 When a check sum error is detected in a received 8 packet or when the BRT word count remaining in a data 9 block reaches zero as a packet is stored into memory, 1~ the BUS RECEIVE microprogram sets the X bus or Y bus 11 completion interrupt flag. O~herwise, the microprogram 12 issues the RINT signal (see Fig. 8) to the inqueue - 13 Processor Empty state logic 101 to allow another packet 14 to be received. When the completion flag i5 set, however, the RINT signal i5 not issued.

17 It is thus the responsibility of the bus receive 18 completion software interrupt handler to issue the RINT
g signal ~by means of an RIR soft~Jare instruction) to reenable the inqueue 65. Until this occurs, the inqueue Bus Fill 21 state logic 93 remains in the FULL state and no additional 22 packets will be received, 24 The completion interrupt signal can thereore 25 designate either a block data transfer that has been sent 2~ and received without exror, or it can designate a partial 27 transfer in which a check sum error is detected, and in 28 which partial transfer of the completion interrupt is 29 generated as a result of the check sum error detected.
30 In the latter case, the sender continues to send the data ~3672~3 1 block but the receiver discards the data block after ~h~
2 check sum error has been detected~ This error shows up i in the bus receive table (BRT) count word as a negative 4 value. This will become more apparent from the description of the operation which follows.

7 The SEND instruction is an instruction that ~ requires four parameter words in the CPU register stack.

The first of the four parameter words is a 11 count or the number of words to be transferred. This value 12 must match the number expected by the BR~ in the receiver 13 processor module if the transfer is to complete success-14 fully.
1~ The second parameter word is the address, minus 17 one, in the System Data area in the sender processor's 18 memory where the data to be transferred is located.
19 ' The third parameter word is a timeout value 21 allotted to completing a single packet ~fifteen data word) 22 transfer. The timeout period is restarted for each packet 23 transferred by the SEND instruction, "` 24 The fourth parameter word specifies the bus ~` 2~ (whether the X bus or ~he Y bus) to be used and speci~ies 27 the receiver processor module. The high order bit of the 2.8 parameter specifies the bus and the low order four bits, 29 in one specific implementation o the invention, specify ~ 30 the number of the receiver processor module, `i` 6~

' ' ~.
' 113~728 1 At the completion of a SEND instruction, there 2 are two possible conditions.

4 The first condition is that a packet timeout occurred and the remaining packets were no~ transmitted 6 and the instructio. was terminated at that point. In 7 this event the remaining pac~ets of the block are not 8 transmitted~

The second condition is an indication that a 11 successful data block transfer has been completed.

13 Thus, in initial summary of the SEND operation, i4 the SEND instruction fills the outqueue buffer 69 (Fig. 4) 15 with fifteen data words, appends an odd-parity check sum, 16 and signals the bus controller 37 that it has a packet 17 ready for transmission~ After each sixteen word packet 18 is tranrsmitted, execution of the SEND instruction resumes 19 at the point where it left off. If the last packet of 20 the block has less than fifteen words, the remaining words 21 are filled in with zeros. The instruction terminates when 22 the last packet is transmitted.
` ~3 24 Fig. S shows the logic diayram and Fig. 7 shows 25 the state diagram for the send hardware.

27 The first action of the SEMD instruction 28 sequence is to issue the S/RECEIVE signal to the processor 29 fill state logic 73 (Fig. 4) and to suppl~ on the M Bus 30 (Fig. 4) the receiver processor number to the receive .

672~

1 register 71. Simultaneousl~, the pointer o the outqueue 2 pointer 79 is set in accordance with the high order bit 3 of the M Bus to connect the outqueue 67 to either the X
4 bus or the Y bus.

~ The store receive (S/RECEIVE) signal causes the 7 processor fiil state logic 73 (which is initially in the 8 empty state as shown in Fig. 7) to advance to the FI~L
9 state as shown in Fig. 7. This state transition causes the receive register 71 (Fig. 4) to be loaded with the ll -receiver processor number.
1~ , 13 At this point the outqueue section 67 is ready 14 for the data packet to be loaded into the outqueue buffer 69. Now, up to fifteen words are read from memory and 16 are stored, by means of the M bus (Fig. 4), into the out-17 queue buffer 69. The store outqueue signal causes each 18 word on the M bus to be written into the outqueue buffer 19 69 in a location specified by the outqueue counter 77.

Each store outqueue signal also causes the outqueue counter 21 77 to be advanced by one.

23 As the words are being read from memory, the 24 address word is being incremented by one, and the count .
of the words to be sent is being decremented by one.
26 If the count reaches zero before fifteen words are read 27 from memory, the rernainder of the outqueue buffer is 28 filled with zeros to pad out the data packet.

. . , ; , 6~
' i~36 728 1 In addition, as the ~ords are bei~g lo~ded into `~ 2 the outqueue buffer 69, ~he microprogram 11~ (Fig. 2) is 3 calculating a modulo-two sum of the data words. Af~er 4 the fifteenth data word has been loaded, this odd check-sum word is loaded into the si~t~enth location of the 6 outqueue buffer 69.
8 At this time the outqueue count~r 77 has a value 9 of count 15 and this value, in combination wi~h the store outqueue signal, causes the processor fill state logic 73 11 to advance from the FILL state to the FULL state as shown 12 in Fig. 7.

14 At this point the microprogram 115 has completed loading of the data into the outqueue 69. The microprogram 16 now waits for the packet to be transmitted by testing for 17 occurrence of the ready (RDY) signal shown in Fig. 7.

19 While waiting for the packet to be transmitted, the microprogram 115 increments a timer; and if the timer 21 runs out or expires before the ready (RDY) signal is 22 -asserted, the microprogram issues the clear outqueue .
"~ 23 ~CLOQ) signal to the processor ill state logic 73 (see ~`` 24 Fig. 4). This causes the processor ~ill state logic 73 ,-~ !
25 to return to the empty state as shown in Fig. 7, and the 26 microprogram then texminates the SEND instruction with 27 the time ou~ indication.

29 In normal operation, the FULL state o~ the 30 processor fill state logic 73 quali~ies the bus empty 1 ~

~3~:i72~

1 state logic 75 to advance from the IDLE state to thc SYNC
2 state shown in Fig. 7. Next, the SYNC state automatically 3 advances to the SEND state, and this state causes the 4 SEND REQUEST signal to be issued to the bus controller 37.
S The SEND REQUEST signal initiates a packet t~ansfer 6 sequence describe~ earlier.

8 As described earlier, when the sender processor 9 module has been identified by the bus controller 37 by 10 polling, and when the receiver processor module has 11 accepted the packet transfer by means of the ~ECEIVE
2 ACKNOWLEDGE signal, the data packet is gated from the 13 outqueue buffer 69 through the outqueue pointer 79 to 1~ one of the data buses 57 for loading into the inqueue of 15 the receiver processor module.

17 As the sixteenth word is gated to the bus, the 18 value of the outqueue counter count 15, in combination with 19 the SEND CO~ND signal and the SE~DER SELECT signal causes 20 the SEND state of the bus empty state logic 75 to advance 21 to the DONE state.

23 The DONE state qualifies the FULL state o~ the 24 processor fill state logic 73 (as shown by the dashed line 25 arrow goiny from the DONE state to the indicated transition 26 from the FULL state in Fig. 7J to advance to the WAIT state.
~7 28 Next, the WAIT state qualifies the DONE state 29 to advance to the IDLE state as illustrated by the state 30 diagram in Fig. 7.

.

,7;;~

1 Finally, the IDLE state quali~ies the W~IT state 2 to advance to the EMPTY state as also indicated in the 3 state diagram of Fig. 7.

The EMPTY state, of the processor fill state 6 logic 73, provides ~he READY indication to the micro-7 program 115.

9 If the packet just transmitted was the last 10 packet in the specified data block, the SEND instruction 11 is terminated and the successful block transfer indication 12 is gi~en 14 If the packet transmitted is not the last 15 packet in a data block, then the se~uence described above 16 is repeated until all words in the block haue been trans-17 mitted, ox until a timeout error has occurred.

19 The SEND instruction is interruptable and 20 resumable; however, the SEND instruction is only interrupt-21 able ~etween packets; and the interruption of the SEND

22 instruction has no effect on the data transmitted.

24 ~hus, by means of a single software instruction tthe SEND instruction) a data block o up to 32,767 words 26 is transmittable from a sender processor module to a 27 receiver processor module, and accuracy of the transmission ~- i 28 is checked by the packet check-sum. Also, the trans-29 mission occurs at a high data transfer rate, because 30 the buffering provided by the outqueue buffer 69 of the 72~

1 sender processor module enables the trans~er to he made 2 at interprocessor bus speed ind~pendent of the memory 3 speed of the sender proc~ssor module. This allows efficient 4 use of this communication path ~etween a number of pro-5 cessor modules on a time slicing basis.

7 As noted above, there is no instruction for 8 receive.

0 For a processor module to receive data over 11 an interprocessor ~us, the operating system in that pro~
12 cessor module must first configure an entry in the bus 13 receive table (BR~). Each BRT entry contains the address 14 where the incoming data is stored and the number of 15 words expected.

17 While the sender processor module is executing 18 the sen~ instruction and sending data over the bus, the l9 bus receive hardware and the microprogram 115 in the 20 receiver processor module are storing the data away 21 according to the appropriate BRT entry tthis occurs inter-22 leaved with software program execution).

24 When the receiver processor module receives 25 the expected number of words from a given sender, the 26 currently executing program i5 interrupted, and that 27 particular bus transfer is completed.

29 Fig. 5 shows the logic diagram and Fig. a shows 30 the state diagram for the bus receive hardware.

~ ~ ~9~
I ~ ~ ~ ~ 4~ ~

l As previously pointed out, there are identical 2 X and Y inqueue sections 65 in ~ach proc~ssor modul~ for the 3 X bus and the Y ~us. Only one of the inqueue s~ction~

4 will therefore be referred to the description which follo~ls.
S
6 After inltial xeset of a processor module, or 7 after a previous receive operation, the RESET state o~ the 8 processor empty state logic 101 advances to the READY
9 state. The READY state qualifies the SYNC state of the bus 10 fill state lo~ic 93 to advance the logic to the ACKNOWLEDGE
11 state.

13 In this ACKNOWLEDGE state the inqueue section 65 ~4 returns RECEIVE ACKNOWLEDGE to the bus controller 37 in 15 response to a SELECT 63 (see Fig. 2) of that processor 16 module 33. This indicates the readiness of the X inqueue 17 section 65 to receive the data packet.

19 In the packet transfer sequence ~described in 20 detail above) the combination of the SELECT of that 21 processor module and the RECEIVE COMMAND signal qualify 22 the ACKNOWLEDGE state of the bus fill state logic 93 and 23 to advance to the RECEIVE state.

At this state transition the sender register 95 26 (Fig. 5) is loaded with the number of the sending processor 27 module.

29 In the R~CEIVE state the data packet is loaded 30 from the data bus to the inqueue buffer 97 under control 31 of the inqueue counter 99.

.~ .
: i ~3~

1 As the sixteenth word of the pacie~ is loaded, 2 it causes the RECEIVE state to advance to the ~ULL state 3 (see Fig. 8).

, Now the FULL state qualifies the READY state 6 of the processor e~pty state logic 101 to advance to 7 the MICROINT~RRUPT state as shown in ~iy. ~. The MICRo-8 INTERRUPT state presents an INQ~E~J~ FULL state to the 9 CPU in~errupt logic. This INQUEUE FULL signal causes a 10 microinterrupt to occur at the end of the next so~tware 11 instruction if the ~L~SK bit corresponding to that bus is on.

13 The bus receive microprogram 115 activated 14 by the interrupt first of all issues a LOCK signal (see 15 Fig. 5) to the processor empty state logic 101. This 16 causes the MI~ROINTERRUPT state of the processor empty 17 state logic 101 to advance to the DUMP state.

19 The LOCK signal also selects either the X
20 inqueue or the Y inqueue; subject, howe~er, to the 21 condition if both inqueues are full and enabled, the X
22 queue is selected.

24 Next, the microprogram 115 issues the K/SEND
25 signal which causes the sender register 95 contents to be 26 gated to the K bus (as shown in Fig. 5) to obtain the 27 packet senderls processor number.

29 Using this processor number, the microprogram 30 115 reads the sender processor's BRT entry to obtain the 31 address and count words.

~ -~36~7~8 1 If the count word is zero or neyative, the 2 packet is discarded; a~d in this c~5e, the microprogram 3 115 issues a RINT si~nal which causes the processor 4 empty state lo~ic;101 to advance from the DUMP state to 5 the RESET state as shown in Fig. 8. In this event there 6 is no further acticn. The microinterrupt is terminated, 7 and software instruction processing is resumed.

9 If the count is positive, the microprogram 115 10 reads words from the inqueue buffer 97 to the K bus by 1~ means of the X/INQUEUE signal as shown in Fig. 5.

13 With each occurr~nce of the K/INQUEUE signal, 14 the inqueue counter 99 is incremented to scan through the 15 inqueue buffer 97.

17 As each data word is read ~rom the inqueue 18 buffer 97, the count word is decremented, the memory 19 ad~ress word is incremented, and the data word is stored 20 into memory.

2~ If the count word reaches zero, no more words 23 are stored in memory, a cornpletion interrupt flag is set 24 and the sender processor number is saved in a memory 2S location. In that event the fill state bus logic 93 stays 2~ in the FULL state until cleared by a software RIR instruction.

28 Thus, when a data block has been completely 29 received, the count word will contain a value between minus 30 14 and zero. After the complekion interrupt occurs, no 1~3.~;7Z~

1 further transfers to the processor over ~he bus ~lhich 2 cause the interrupt are permitted until the inqueue is 3 cleared with an RIR instruction.

S ~s the data words are stored into the memor~l, 6 a modulo-two sum o~ packet data is calculated.

8 If the check sum is bad, the word count in the g BRT entry is set to minus 256, a completion interrupt flag 10 is set, and the sender processor number is saved in memory.
11 As above, the bus fill state logic 93 stays in the FULL
12 state until cleared by an RIR instruction.

~4 If the count word does not reach zero, and the 15 check sum is good, the bus receive microprogram 115 issues 16 the RINT signal to the processor empty state logic as shown 17 in Fig. 5 which causes the DUL~P state of ~he processor 18 empty state logic 101 to advance to the RESET state as 19 shown in Fig. 8.
21 The RESET state of the logic 101 qualifies the 22 bus fill state logic 93 to advance from the FULL state to 23 the SYNC state as also shown in Fig. 8.

At this point, the logic hag been returned to 26 the state it was in beore the packet was received, thus 27 enabling the receipt o~ more packets.

~3~

1 These packets may be from the same send~r, 2 completing that data block, or the packets ma~I be from 3 some o~her ~ender.

This completes the action of the bus receive 6 microprogram 115 a~d the microprocessor 113 resumes pro-7 cessing of software instructions.

g When a bus receive completion interrupt has 10 occurred, the software interrupt handler obtains the sender 11 processor number from the memory location where that number 12 was saved~ and the so~tware interrupt handler can then 13 detect if a check sum error occurred by examining that 14 sender processor's bus receive table count word.

16 In the case of a transmission error, the count 17 word has been set to minus 256. Otherwise, the count word 1~ will contain a value between minus fourteen and zero.

As mentioned above, it is thus the responsibility 21 Of the bus receive completion software interrupt handler 22 to issue the RINT signal (by means of an RIR software 23 instruction) to reenahle the inqueue 65.

In summary on the receive operation, just as 26 the sending of a data block by a ~ender processor module 27 is viewed by software as a single event, the receiving of 28 data by a receiver processor does not cause a software 2g interrupt of the receiver processor module until the 30 entire data block has been received or until an error has ~3~2~3 1 has occurred. Also, the inqu~ue~ 65 s~r~e as ~uffers to 2 allow the transmission of data to occur at bus transmission 3 rates while allowing the storing of data into memory and 4 the checking of the data to occur at memory speed. This 5 ability to use the high transmisslon rate on thc bus insures 6 adequate bus bandwidth to service a num~er of processor modules 7 on a time slicing basis. Finally, ~he provision of a check 8 sum word in each data packet provides a means in the receiver g processor module for checking the accuracy of the data 10 received over the multiprocessor communication path.

12 Xnformation sent over the interprocessor bus is 13 sent under the control of the operating system and is sent 14 from one process in one processor module 33 to another pro-15 cess in another processor module 33. A process (as described 16 in detail above in the description of the Multiprocessor 17 System) is a fundamental entity of control in the software 18 system; and a number of processes coexist in a processor l9 module 33. The information sent over the interprocessor bus 20 between processes in different processor modules consists 2L Of two types of elements, control packets and data.

23 The controL packets are used to inform the 24 receiving processor module 33 about message initiations, 25 cancellations, and data trans~ers.

27 In this regard it should be noted that, while 28 the interprocessor ~uses 35 interconnect the processor 29 modules 33, a process within a particular processor 30 module 33 communicates with another process or with 31 other processes within another processor moduLe 33 through 32 a method of multiplexing the interprocessor bus 35. The 7~

~6~

1 bus traffic between two processor modules 33 ~ill there-2 fore contain pieces of interprocess co~munications 3 that are in various states of compl~tion. ~any inter~
4 process communications are therefore b~iny interle~ved 5 on an apparently simultaneous basis.

7 The hardware is time slicing the use of the 8 interprocessor bus 35 on a packet level, and multiple 9 processes are intercommunicating both within the pro-lO cessor modules 33 and to the exten~ necessary over the 11 interprocessor buses 35 in message transactions which 12 occur interleaved with each other. Under no circumstances 13 is an interprocessor bus 35 allocated to any specific 14 process-to-process co~munication.

16 Data informa~ion is sent over the interprocessor 17 bus in one or more packets and is always preceded by a 18 control packet and is always followed by a trailer packet.

The control packet preceding the data packets 21 is needed because a bus is never dedicated to a specific 22 message, and the control packet is therefore needed to 23 correctly identify the messaye and to indicate how much ~ .~
24 data is to be received in the message.

26 This information transer (control packet, data 27 information, trailer packet) is made as an indivisible 28 unit once it is started. The sender processor module 29 sends the data block as an individual transmission (con-30 sisting of some number of data packets) and sends the 7g 3~;~2~

1 trailer packet as an indiviclual transmission; and onl~
2 then is the sender proc~ssor module able to send 3 information relating to another message.

The trailer packet serves two purposes.

7 First of all, if there is an error during a data 8 transmission (and therefore the rest of the data block must 9 be discarded), the trailer packet indicates the end of ¦ 10 the block.

12 Secondly, if the sender attempts ~o send too 13 much data (and again the block must be di~carded), the 14 tr~iler packet provides a means for recognizing data has 15 been transmitted and the data transmission has completed.

17 The information transmitted is either duplicated 18 over different paths (so that it is insured that the 19 information will get to the receiver) or a receiver acknowledg-20 ment is required (so that the information is repeated if ;21 necessary). Any single bus error therefore cannot cause 22 info~mation to be lost, and any single bus error will not 23 be seen by the two processes involved.

~he bus receive software interlocks with the 26 bus receive hardware (the inqueue section 65 shown in 27 Fig. 2) by controlling the trans~er of information from 28 the inqueue into the memory 107, i .~ I

~ . .
~ 80 . .

~3~7:2~

l This allows such operations as changing the 2 bus receive table information to be done without race 3 conditions ts~nchronization problems).

Once the bus receive t~ble information has 6 been updated the interlock ls removed by cleariny the 7 previous completion interrup-t and by reenabling the 8 bus receive microinterxupts by setting on the bus mask 9 bit in the mask register.

11 This does two things. It allows the inqueue 12 hardware to accept a packet into the inqueue, and it also 13 enables the bus recei~e microprogram to transfer that 14 inrormation from the inqueue into memory.
16 The hardware/software system is so constructed 17 that no information is lost on a system power failure (such 18 as a complete failure of AC power from the mains) or on a l9 line transient that causes a momentary power failure for 20 part of the system.

22 This hardware/software system coaction includes 23 a power warn signal (see line 337 of Fig. 3~ supplied to 24 the inqueue section 65 (see Fig. 2) so that, at most, one 25 further packet of information can be loaded into the in-26 queue after the receipt of the power warn signal.
., .

`~ 28 The sotware action in this event includes a `;~ 29 SEND instruction to force the inqueues to be full. The ~ 30 ~et effect is to insure that no transmissions are completed ,~ -~1 .

~ ~367;~

1 after the proce~sor module 33 has received its power 2 warn signal, so that the state of every transer is 3 known when logic power is removed.
The interprocessor buses 35 are used b~ the 6 operating system to ascertain that other processor 7 modules in the system are operating. Every N seconds, 8 each of the processor modules 33 sends a control packet 9 to each processor module 33 in the system on each 10 interprocessor bus 35. Every two N seconds, each pxo-11 cessor module 33 must have received such a packet from 12 each processor module 33 in the system. A processor 13 module that does not respond is considered down. If a 14 pr-cessor module does not get its own message, then that 15 processor module 33 knows that something is wrong with 16 it, and it will no longer take over I/O device controllers 17 41.

19 Fig. 42 diagrammatically illustrates how a 20 particular application program can run continuously even 21 though various parts of the multiprocessor system can ~22 become inoperative.

;24 Each of the separate views shown in Fig. 42 25 illustrates a multiprocessor system con~iguration which 26 eonsists of two proeessor modules 33 connected by dual 27 interprocessor buses 35 (indicated as an X bus and a Y bus), .-:
28 a device controller 41 which controls a number of keyboard 29 terminals, and another device controller 41 which controls -i30 a disc.
, .

The indlvidual views of Fig. 42 indicate various 2 parts of the multiprocessor system rendered unserviceable 3 and then reintroduced into the multiprocessor system 4 in a serviceable state.

6 The sequence starts with the upper left hand 7 view and then proceeds in the order indicated by the 8 broad line arrows between the views. The sequence thus g goes from the condition indicated as (l) Initial State lO to (2) CPU 0 Down to (3) CPU 0 Restored to (4) CPU l Down ll to ~5) CPU l Restored (as indicated by the legends above 12 each individual view).

l~ In the initial state of the multiprocessor system 15 shown in the view entitled "Initial State" at the upper 16 left hand corner of Fig. 42, one copy (PA) of the application `17 program is active. This copy makes a system call to create 18 the copy PB as a backup to which the application program 19 PA then passes information. All of the I/O is taking 20 place by way of the processor module 0. In this initial 21 state either interprocessor bus 35 may fail or be brought 22 down (as indicated by the bars on the X bus) and can be ~3 then reintroduced into the multiprocessor system without ~`
24 producing any effect on the application program PA.

; 26 ,: .

~8 ~;

~S7Z8 In the next view (the view entitled "CPU 0 Down") the processor module 0 is rendered unserviceable.
The multiprocessor system informs the applica-tion pro-gram PA that this has happened, and the application program PA no longer -tries to communicate with the pro-gram Ps. All of the I/O is switched by the multiprocessor system to take place by way of the processor module 1, and the application program continues to service the terminals without interruption over the I/O bus 39 connect-ing the processor module 1 with the device controllers 41 (as indicated by the solid line arrow on the right hand I/O bus 39).
In the next state of operation of the multi-processor system, as illustrated in the center top view of Figure 42 and entitled "CPU 0 Restored", the processor module 0 is now brought back into service by way of a con-sole command. The processor module 0 is reloaded with the multiprocessor system from the disc by way of the processor module 1. The application program PA is informed that processor module 0 is now serviceable and the application program PA tells the multiprocessor system to create another copy oE the application program in the processor module 0.
This other copy is desiynated as PC. The terminals con-tinue in operation without interruption.

~ 84 -~z~

Next, the processor module 1 is rendered in-operative, as illustrated in the view entitled "CPU 1 Down". ~he application program PC is informed of this fact by the multiprocessor system an~ the application pro-gram PC takes over the applica-tion. The multiprocessor system automatically performs all of the I/O by way of the processor module 0. The terminals continue without inter-ruption.
Finally, as indicated by the top riyht hand view of Figure 42 entitled "CPU 1 Restored", the processor module 1 is rendered operable by way of a console command and is reloaded with the multiprocessor system from the disc by way of the processor module 0. The application program PC is informed that the processor module is now available, and it tells the multiprocessor system to cxeate another copy of itself (application program PD) in the processor module 1. All elements of the multiprocessor system are now operable.
During the whole of this time both interprocessor buses and both processor modules had been rendered unservice-able and reintroduced into the system, but the application program and the terminals continued without a break.

~31~7Z~

It is an importan-t fea-ture of the multi-processor system that not only can the application program continue while something has failed, but also that the failed component can be repaired and/or replaced while the appllcation program continues. This is true not only for the processor modules and inter-processor buses but also for all elements of the multiprocessor system, such as power supplies, fans in the rack, etc. The multiprocessor system 31 thus is a true continuously operating system.

- ~6 -1~ ;72~

1 THE INPUT/OUTPUT SYSTE~ AND DU~L PO~.T DE~VICE CONTROLLER:

. .i 3 The multiprocessor system 31 shown in Fig. l 4 includes an input/output (I/0) system and dual'port 5 device controIlers 41 as noted generally above.

7 - The general purpose of the I/O system is to 8 allow transfer of data hetween a processor module 33 and g peripheral devices.

11 It is an impoxtant feature of the present' 12 invention that the data transfer can be accomplished over 13 redundant paths to insure fail soft operations so that 14 a failure of a processor module 33 or a failure of a l5 part of a device controller 41 will not inhibit transfer 16 Of data to and from a particular peripheral device.

18 Each device controller 41 has dual ports 43 19 and related structure which, in association with two 20 related I/O buses 39, permi~ the redundant access to a 21 peripheral device as will be described in more detail below.

23 The I/O system of the present invention also 24 has some particularly signi~icant features in terms o 25 performance. For example, one o~ the performance ~26 features of the I/O sy~tem o~ the present invention is 27 the speed (bandwidth) at which,the input/output bus 28 structure operates. The device controllers 41 collect `;~
~` 29 data from peripheral devices which transmit data at . . .
30 relatively slow rates and transmit the collected data ~' .

1 to the processor modules in a burst multiplex mode a~c 2 or near memory speed of the processor modules 33.

As illustrated in Fig. 1, each processor 5 module 33 is attached to and handles a plurali-ty of 6 individual device controllers 41; and this fact makes 7 it possible for each device controller 41 to be 8 connected (through dual ports 43) to more than one g processor module 33 in a single multiprocessor system.

11 With reference now to Fig~ 12 of the drawings, ~2 each processor module 33 includes, in addition to the 13 interprocessor control 55 noted above, a central processor 14 un'.t ~CPU) part 105, a memory part 107 and an input/output ~I/O) channel part 109.

As illustrated in Fig. 12 and also in Fig. 1, 1~ each device controller 41 controls one or more devices 19 through connecting lines 111 connected in a star pattern, 20 i.e. each device independently connected to the device 21 Cntroller.

23 In Fiy. 12 a disc drive 45 is connected to one 24 device controller 41 and a tape drive 49 is connected to - 25 another device controller 41.

27 With continued reference to Fig. 12, each CPU
28 part 105 includes a microprocessor 113. A microprogram 29 115 is associated with each microprocessor 113. A part 30 of the microprogram 115 is executed by the microprocessor 1~iL36~2~

1 113 in performing I/O instructions for the I/O s~stem.
2 The l/O instructions are indicated in Fig, 12 clS EIO
3 ~execute I/0), II0 (interrogate I/0), HII0 (interroyate 4 high priority I/O~; and these instructions are 5 illustrated and described in greater detail below with 6 reference to Figs. 15, 16 and 17.

8 The microprocessor 113 has access to the I/O
9 bus 39 by way of the I/O channel 109 by a collection lO.of paths 117 as illustrated in Fig. 12.

12 With continued reference to Fig. 12, the I/0 13 channel lOg includes a microprocessor 119, and a micro-14 program 121 is associated with the microprocessor 119.
16 The microprogxam 121 has a single unction 17 in the multiprocessor system, and that function is to 18 perform the reconnect and data transfer sequence 19 illustrated in Fig. 16 (and described in more detail 20 below).

22 The I/O channel 109 of a processor module 33 23 also includes (as shown in Fig. 12) data path logic 123.

As best illustrated in Fig. 13, the data path 26 logic 123 includes a channel memory data register 125, 27 an input/output data register 127, a channel memory 28 address register 129, a character count register 131, 29 an active device address reyister 133, a priority resolv-30 ing register 135 and parity generation and check logic 137.

ZI~

1 The path 117 shown in Fig. 12 includes two buses 2 indicated as the M bus and the K bus in Fig. 13.

4 The M bus is an outbus from the microprocessor 113 and transmits data into the input/output data 6 register 127.

8 The K bus is an inbus which transmits data g from the data path logic 123 into the microprocessor 113.

11 With reference to Fig. 12, a path 139 connects 12 the data path logic 123 and the memory subsystem 107.

14 - This path 139 is illustrated in Fig. 12 as including both a hardware path 139A and two logical paths 16 139B and 139C in the memory subsystem 107 of a processor 17 mOdule 33, 19 Logical paths 139B and 139C will be descri~ed 20 in greater detail below in connection with the 21 description o~ Fig. 16.

23 The hardware path 139A includes three branches 24 as illustrated in Fig. 13.

26 A first branch 139A-1 transmits from memory 27 into the channel memory data register 125.

29 A second path 139A-2 transmits from the channel 30 memory address re~ister 129 to memory.
.
so 72~

1 And a third path 139A-3 transmits from the 2 input/output data register 127 to memor~.

4 With reference to Fig. 12, the input/output 5 channel of a processor module 33 includes a control logic 6 section 141.

8 This control logic section 141 in turn includes 9 a T bus machine 143 (see Fig. 13) and request lines 10 RECONNECT IN (RCI) 145, LOW PRIORITY INTERRUPT REQUEST
11 (LIRQ) 147, HIGH PRIORITY INTERRUPT REQUEST (HIRQ) 149 12 and RANK 151 tsee Fig . 14 ) .

14 - The I/O bus 39 shown in Fig. 14 and Fig. 12 15 also includes a group of channel function lines 153, 157 16 and 159. See also Fig. 13~ The TAG bus tT bus) 153 17 consists of four lines which serve as function lines, and 18 there are three lines SERVICE OUT (SVO) 155, SERVICE IN
19 tsvI~ 157, and S~OP IN (STI) 159 which serve as handshake 20 lines as indicated by the legends in Fig. 14.

22 As shown in Fig. 14 and Fig. 12, the I/O bus 2339 also includes a group of data lines 161, 163, 165, 167 24and 169.

26 ~he DATA BUS lines 16$ and PARITY 163 are bi ; 27directional and serve a~ data lines and as indicated in 28Fig. 14, there are sixteen DA~A BUS lines 161 and one 29PARIT~ $ine 163 in this group.

67;~

The lines RND OF TRANSFER (EOT) 165, PAD OUT
(PADO~ 167 and PAD IN (PADI) 169 serve as data status lines, and indicate special conditions that may occur on the data lines 161 and 163 from time-to-time.
Finally, the I/O bus 39 includes a reset line (IORST) 171 as also shown in Figure 14 and in Fiy-ure 12.
Each T bus command illustrated in Figure 18 requires some specific format on the data bus 161 while a T bus command is valid. This specific data bus format is illustxated for the T bus functions Load Address and Command (LAC) and Read Device Status (RDST) shown in Figure 18, for the preferred embodiment.
In the case of the T bus function LAC, the data or field transmitted on lines 0 to 5 of the data bus 161 specify the operation to be performed; the field transmitted on lines 8 to 12 of the data bus specify the device controller 41 (or more precisely the port 43 of that device controller which is attached to the data bus 161~ to which the command is addressed; and the field transmitted on data bus lines 13 to 15 specify which device attached to the device controller is to be operated on by that device controller 41 in response to this command.
In the case of the T bus function RDST, data bus bits ~, 1, 2 and 3 indicate ownership error, interrupt pending, device busy, and parity error respectively. Bits 4 to 15 return device dependent status.

~ 92 -X

~3~ 8 1 Th~ functions on the T ~us ar~ tra~smitt~d 2 in three sequences, shown in Figs. 15, 16 and 17 and 3 described in detail below~

Each T bus function is asserted b~ the channel 6 and a handshake sequence is performed bet~leen the channel 7 109 and the device controller 41 using the handshake lines 8 155, 157 and 159 to acknowledye receipt of the T bus 9 function. Control of the T bus and handshake is the 10 function of the T bus machine 143 in Fig. 13.

lZ Fig. 28 is a timing diagram showing the operation 13 Of the handshake bet~een the I/O channel 109 and the ports 43.

As illustrated in Fig. 28, line 155 transmits 16 the service out signal (SVO) and line 157 transmits the 17 service in signal (SVI).

19 The channel clock cycle is shown in vertical 20 orientation with the SV0 and SVI signals.

22 As illustrated in ~ig. 28, the service in 23 ~SVI) signal is not synchronized with the channel clock 24 and may be asserted at any time by the device controller 2~ in response to a service out signal from the I/O channel 2~ los.

28 Before asserting service out (SV0), the channel 29 109 asserts the ~ bus function and, if required, the data 30 bus.

7Z~

1 The channel then asserts a servicP out signal 2 as indicated by the vertical rise 279 in Fig. ~8; and, 3 SVO remains true until the device controller responds 4 with service in (SVI) (281), acknowledging the channel 5 command; SVI remains true until the channel drops SVO.

7 When the device controller 41 asserts the 8 service in (SVI) signal, the channel 109 removes the 9 service out (SVO) signal (as shown by the vertical drop 10 283 in Fig. 28) in a time period typically between one 11 and two clock cycles; and in response, the device controller 12 drops service in (S~I) as shown by the vertical drop 13 285 in Fig. 28.

When the device controller drops the service 16 in (SVI) signal, the channel 109 is free to reassert a 17 service out signal (SVO) for the next transfer; however, 18 the channel will not reassért SVO until SVI has been 19 droppedO

21 The arrows 281A, 283A and 285A in Fig. 28 2~ indicate the responses to the actions 279, 281, 283 23 respectively.

The handshake is completed at the trailing 26 edge of the vertical drop 285 as shown in Fig. 28, 28 - On an output trans~er, the inter~ace data register 29 213 of the contxoller accepts the data at the leading edge 30 Of service out (vertical rise 279) and trans~ers the data , g4 f ~--Z~ ~

1 to the control part of the device controller lB7 at the 2 trailing edge of ~he service out (the vertical drop 283).

4 On an input transfer the channel 109 accepts 5 data from the device controller at the trailing edge of 6 service out (the vertical drop 283).
8 Thus, a two line handshake i5 used to interloc~
9 transfer of information between the channel 109 and its 10 device controller 41, since they act asynchronously.

12 This is the general handshake condition, 13 indicated as handshake 2L in Figs. 15, 16 and 17.

In addition, two special handshake considerations 16 occur, when approprizte.

:18 . First, channel commands used to select a 19 device controller are not handshaken by SVI, since no 20 single device controller is selected during this time.

22 These commands include (as shown in Fig. 18~:
23 SEL - Select;
24 LAC - Load Address & Command;
2S HPOL - Hi Prioxiky Interrupt Poll;
.26 LPOL - Lo Priorit~ Interrupt Poll; and 27 RPOL - Reconnect Interrupt Poll.

29 Also, commands used to terminate a sequence are 30 not handshaken by SVI since they cause a selected device 31 controller to deselect itself.

~367~15 1 These commands include (as also sho~m in Fi~l. 18):
2 DSEL - De-Select;
3 ABTI - Abort Instruction (I/O), and 4 AsTD - Abort Data.

6 For all of the co~nands noted above which are 7 not handshaken, the channel asserts SVO.(155) for ~ giv~n 8 period of time (e.g., two clock cycles) and then the 9 channel removes SVO. This type of handshake is referred 10 to as Handshake lL in Figs. 15, 16 and 17.
11 `
12 Second, data transfer is handshaken normally 13 except that when a device controller wishes to signal that 14 it does not require further service, it ret~lrns stop-in (STI) instead of SVI. When SVo is next dropped by the 16 channel, the port deselects itself~ STI otherwise hand-17 shakes in the same manner as SVI.

19 A~ a fuxther condition on all handshakes, when 20 the channel prepares to assert SVO, it initiates a timer 21 ~part of T bus machine 143 in Fig. 13) which times out 22 and posts an error if the next handshake cycle is not 23 initiated and completed within the period of time set 24 by the timer. If the timer times out, an error is 25 posted at the appropriate point in the sequence, and 26 either ABTI (EIO, IIO or ~IIO sequence~ or ~BTD
27 (reconnect sequence) is sent to the device controller 41 28 (see discussions o~ Figs. 15, 16 and 17).

3~

tf ~3~7;Z8 1 Fig. 29 shows the logic for the handshake sho~/1n 2 in Fig. 28~ The logic shown in Fig. 29 is part o the 3 T bus machine 143 shown in Fig. 13. The logic shown in 4 Fig. 29 is the logic which is ef~Pctive for the general 5 handshake condition noted above.

7 The logic shown in Fig. 29 includes a service 8 out flip-flop 287 and a service in synchronization flip-9 flop 289. As illustrated by the dividing lines and 10 legends in Fig. 29, the flip-flops 287 and 289 are 11 physically located within the channei 109.
~2 13 The devlce controller 41 includes combinational 14 logic 291 and a transmitter 293 which transmits a service 15 in signal ~SVI) back to the D input of the flip-flop 289.

17 The ~u~ctioning of the logic shown in Fig.
18 29 is as follows.

The channel 109 asserts service out by turning 21 on the J input of the flip-flop 287; and when the next 22 clock cycle starts, the service out signal is transmitted 23 by a transmitter 295 to the device controller.

When the combinational logic 291 in the device 26 controller is ready it enable5 the transmitter 293 to 27 return the service in signal (SVI) to the flip-flop 289.
28 This completes the handshake.

.

~3~72~

1 Turning now to the dual port devic~ controller, 2 as illustrated in Fig. 19, each of the dual ports 43 in 3 a device controller 41 is connected by a physical 4 CQnnectiOn 179 to interface common logic 181 (shown in 5 more detail ln Fig, 21) and each of the ports 43 is ~ also associated through a logical connection 183 to the 7 interface common logic 181 as determined by an ownership 8 latch 185.

1~ As shown by the connecting line 180 in Fig.

11 19, the interface common logic 181 i5 associated with 12 the control part 187 of the device controller 41. The 13 control part 187 of the device controller includes a 14 buffer 189.

16 The dual ports 43 shown in block diagram form 17 in Fig. 19 (and ln more detail in Fig. 23) are important l~ parts of the multiprocessor system of the present 19 invention because the dual ports provide the failsoft 20 capability for the I/0 system.

22 The ports 43 and related system components are 23 structured in such a way that the two ports 43 of one 24 device controller 41 are logically and physically 25 independent of each other. As a result, no component 26 part o one port 43 is also a component of the other 27 port 43 of a particular device controller 41; and no single 28 component failure (such as an integrated circuit failure) 29 in one port can affect the operation of the other port.

f r ~3~7~

1 Each port 43 functions to interface (a5 2 ind~cated by the legend in Fig. 19) a processor module 3 33 with a device con roller, and ultimately with a 4 particular device, through the device controller 41.
5 The port 43 lS the entity that com~unicates with th~
6 processor module and communicates with the control part 7 of the device controller 187 (conditional on the state 8 Of the ownership latch 185).

That is, the port itself makes the connection 11 to a processor module (dependent upon instructions 12 received from the I/O channel 109 as discussed in more 13 detail below) by setting its select bit 173.

Each of the individual ports 43 in a particular 16 device controller 41 can be connected independently to 17 a processor module 33 and at the same time as the other 18 port in that device controller is connected to a different 19 module. However, the ownership latch 185 establishes 20 the logical connection between the control part of the device controller and one of the dual ports 43 so that 22 only one port has control of the devlce controller at any 23 one point in time.

The decode logic determines what function 26 is being transmi~ted on the T bus 153 at any particular time.

28 The control logic combines T bus ~unctions 29 to perform specific port functions, or e~ample, set 30 select bit, clear select bit, read interrupt status.

gg ~L~.3~

1 ~he functioning of the control loyic is 2 illustrated in the logic equations set out ln Fig. 27.

4 When a connection sequence (to be described 5 later in reference to Figs. 15, 16 and 17) is transmitted 6 over the I/O bus 39, one of the ports 43 (and only the 7 one port 43 in a device controller 41 attached to that 8 I/0 bus 3~) connects (in a loglcal sense) to the bus 39 9 by setting i~s select bit 173.

11 This logical connection is determined by part 12 of the data transmitted in that connection sequence.
13 When connected, that particular port 43 subsequently 14 responds to channel protocols in passing information 1~ between the channel and the control part of the de~ice 16 controller. The device address comparator 193 is the 17 component part of the port 43 that determines the port's 18 unique address.

The device address comparator 193 determines 21 the unique address for a particular port 43 by comparing 22 the device address field on the data bus 161 during a 23 LAC T bus function, with device address jumpers associated 24 with a particular port 43. When the address transmitted 25 by the channel 109 matches the address determined by the 26 jumpers on a particular port 43, the term ADDCOMP (see 27 Fig. 27) is generated and the select bit 173 for that 28 port is set (assuming that the other cond~tions set out 29 in Fig. 27 allow the select bit to be set). The port 3~ 43 then responds to all T bus operations until the sequence 31 terminates by clearing the select bit.

_".

The abbreviations used in Figure 27 include the following: !
ADD Comp - Address Compare (Device Addregs);
PAROKFF - Parity OK Flip-Flop;
SEL - Select;
OWN - Ownership; and SELBIT - Select Bit.

.~;
.~ -.' ~

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

~'. ~ "' 101 - .

'`'~ .
' ~ ' , ' ~;.
~'. ~' '- ' . '' : :
X ;, :, .~ ."`

~ r ~1367213 1 The parity check register 177 is related to 2 the parity generator and check logic 137 of Fig. 13 in 3 that on output the parity generator loyic 137 generates 4 the parity to be checked by the parity checker 177 of - 5 the port 43, and this parity must check or the operatlon 6 will be aborted by the I/O channel 109 of the processor 7 module 33. On input, the interface common logic 181 8 generates parity to be checked by the channel parity 9 check logic 137 in a similar fashion.

~0 ~ As shown in Fig. 24, the parity c'neck is 12 started before data is,loaded into the register, and 13 the parity check is continued until after the data has 14 been fully loaded into the register. That is, the 15 parity on the D bus is checked by the port parity 16 register whene~er the channel asserts SVO with an output 17 T bus function, and the parity is monitored for the 18 duration of SVO to insure that the data on the D bus is 19 stable for the duration of SVO while the port transfers 20 the data into the data register 213.

22 This parity check occurs on each transaction 23 in a T bus sequence; and if a parity error occurred during 24 any transaction in the sequence, the error is returned as a status bit in response to a ~ bus function during a 26 sequence. For example, in an EIO sequence tFig~ 18 and 27 15) the P bit return or RDST indicates that the port 28 determined a parity error duriny the EIO sequence.

7Z~

1 As illustrated in Fig. 18, the parit~ error 2 bit is a bit number 3 on the D bus in response to a 3 RDST function on the T bus.
If a parity error occurs at some time o~her 6 than during an EI0 sequence, the parity error is reported 7 during the read interrupt status (RIST) T bus function 8 similar to the manner described above for the RDST T
9 bus function.

11 The parity error is cleared at the beginning ~2 of an EIO, IIO, HIIO or reconnect se~uence as shown 13 in Fig. 24.
1~
If a parity error is detected during any 16 sequence it is recorded by the parity check register 17 to be returned on the D bus in response to a RDST or 18 RIST T bus-function.

With continued reference to Fig. 20, the 21 function of the enable latch 175 in the port 43 is to 22 allow the I/O system to recover rom a certain class of 23 errors that would otherwise render inoperative both of 24 the I/O buses 39 attached to a particular device controller 25 41. The enable latch 175 accomplishes this b~ not allow-26 ing the port 43 to place any signal5 on the I/O bus 39.
~7 28 The enable latch 175 is cleared by a specific 29 disable command. This is a load address and ~ommand (LAC~
30 T bus function with a specific operation code trans-31 mitted on the D bus 161.

_, ~3~

1 Once the ena~l~ latch 175 is clear~d, this 2 enable latch cannot be program~atically reset.
The port 43 includes a status multiplexer 195.
5 The status m~ltiplexer 195 returns the ownership err~r 6 mentioned above if the device controller 41 is logically 7 connected to the other port 4~ of that device controller, 8 to indicate that the device controller is owned by the 9 other port and commands to this port will be ignored.

11 The port 43 includes an interface transceiver 12 137 for each input line (i.e., SVI, STI, Data Bus, Parity, 13 PADI, RCI, LIRQ, HIRQ~ of the I/O bus 39 sho~m in Fig. 14.
14 The transceivers 197 transmit data from the port 43 to 15 the I/O channel 10g when the port select bit 173 is 16 set and the T bus function on the T bus 153 requires 17 that the devic~ controller 41 return information to the 18 channel. The transceivers 197 pass information from the l9 data bus 161 into the port 43 at all times.
21 It is a feature of the present invention that 22 the power on circuit 182 acts in association with the 23 transceivers 197 to control the behavior o~ the txans-24 ceivers as the device controller 41 is powered up or 25 powered down, in a w~y which prevents erroneous signals 26 rom being placed on the I/O bus while power is going up 27 or down. This featuxe ls paxticularly significant from 28 the standpoint of on line maintenance.

1~3~'7ZI!~

1 As shown in Fig. 20, each transceiver 197 2 comprises a receiver 198 and a transmltter 200.
4 The txansmitter is enabled by an enable line 5 202.

7 There are several terms which are on the enable 8 line 202. These include the select bit 173, a required 9 input function on the T bus, and a signal from the PON
10 circuit 182.

12 The signal from the PON circuit, in a particular 13 embodiment of the present invention, is connected in a ~4 "wire or" connection to the output of the gate which 15 combines the other terms so that the output of the PON
16 circuit overrides the other terms by pulling down the 17 enable line 202. This insures that the transmitter 200 18 (in one specific embodiment, an 8T26A or 7438) is placed 19 in a high impedence state until the PON circuit detects 20 that the power is at a sufficient level that the integrated 21 circuits will operate correctly. The PON circuit output 22 stage is designed to take advantage of a property of the 23 spPcific transceiver integrated circuit used~ On this 24 particular type IC if the driver e~able line 202 is held 25 below two diode drops a~ove ground potential, the trans-.26 mitter output transistors are Eorced into the of state 27 regardless of the level of power applied to the integrated .28 circuit. This ensures that the driver cannot drive the bus.
29 . .

~;,.

~L3~;7Z~

This particular combination of fea-tures provides a mode of operation wherein the output of the integrated circuit is controlled as power comes up or goes down, whereas normally the output of an integrated circuit is undefined when power drops below a certain level.
This same circuit is used on the X and Y buses of the interprocessor bus system to control the transceivers and control signals generated by the interprocessor control 55. As indicated in Figure 30, each central proce.ssor unit (CPU) 105 has a PON circuit 182 which is similar to the PON circuit 182 in the device controller. The PON circuits therefore control the transmitters Eor all of the device controllers 41 and all of the interprocessor controls 55.
Details of the power-on (PON) circuit are shown in Figure 25 where the circuit is indicated generally by the reference numeral 182.
The purpose of the PON circuit is to sense two different voltage levels of the five volt supply.
If power is failing, the circuit senses the point at which power drops below a certain level which renders the logic in th~ device controller or CPU an indeterminate state or condition. At this point the circuit supplies signals to protect the system against the logic which subsequently goes into an undeinable state.

:

.~ - 106 -. ~

The second volta~c lcvcl ~ ich the ~'ON circuits will sense is a value that is pcrceived whon power is coming up. This second level at which power is sensed will be greater than the first level by roughly 100 millivolts LO provide hysteresis for the system to eliminate any conditions of oscillation.
The PON circuit stays in a stable condition after it senses one of the voltage conditions until it senses the other voltage condition, at which point it changes state. The state at which the PON circuit is in at any particular time determines ~he voltage level at which the transition to the other state will be made.
The power on circuit 182 thus presents a signal establishing an indication that the power is within predetermined, acceptable operating limits for the device controller 41. If the power is not within those predetermined, acceptable operating limits, the signal output of the power-on circuit 182 is used to directly disable the appropriate bus signals of the device controller 41.
The output of the PON circuit 182 is a binary output. If the output is a one, the power is within satisfactory limits. If the output of the PON circuit is a zero, this is an indication that the power is below the acceptable limit.
.

i~
~L~3~

1 The power-on cixcuit 182 shown in Fig. 25 and 2 to be described in detail below is used with the devic~
3 controller 41 and has seven output driver stages which 4 are used in the application of the power-on circuit 182 5 to the device controller 41. However, the same power-on 6 circuit 182 is also used with the CPU 105 and the bus controller 37, but in those applications ~he power-on 8 circuit will have a lesser number of output driver stayes.

As illustrated in Fig. 25, the PON circuit 182 11 comprises a current source 184 and a differential amplifier 12 186.

14 The dif~erential amplifier 186 has, as one 15 input, a temperature compensated reference voltage input 16 on a line 188 and has a second input on a line 190 which 17 is an indication of the voltage that is to be sensed by 18 the powe.r-on circuit.

The reference voltage on line 188 is established 21 by a zener diode 192.
~2 23 The differential amplifier 186 comprises a 24 matched pair of transisfors 194 and 196.

~: 26 The voltaye applied on the liné 190 i~
27 determined by resistors 198, 200 and 202. The re~istors 28 198, 200 and 202 are metal film reslstors which provide 29 a high degree of temperature stability in the PON
30 circuit.

' f ~L~367 Th~ outputs on lines 204 and 206 of the ~ dif~er2ntial amplifier 186 are applied to a thrce 3 transistor array (the transistors 208, 210 and 212~, 4 and this thr~e transistor array in turn controls the 5 main output control transistor 214.
7 The main output control transistor 214 8 drives ail output drivers that are attached. For example, g in the application of the PON circuit 182 for the de~ice 10 controller 41 (as illustrated in Fig. 25), the main out-11 put transistor 214 drives output stages 216 through 228.
12 The output stage 216 i5 used to clear the logic, the out-13 put stages 218, 220 and 222 are used in combination ~lith 14 the interface devices of one port 43 of.the device controller 15 41, and the output stages 224, 226 and 228 are used in 1~ combination with the interface device of the other port 17 43 of the device controller 41.

1~ Finally, the PON circuit 182 includes a 20 hysteresis control 230. The hysteresis control 230 ~1 includes resistors 232, 234 and a transistor 236.
2~ - _ ~3 In operat.ion, assuming that operation is 24 started from a power off statè to a power on condition, 25 the power is applied through the current source 182 to 26 the differential amplifier 186 and to the main output 27 control transistor 214. At this time the voltage on 28 the line 190 is less.than the voltage on the line 188 29 so the differential amplifier 186 holds the output of 30 the main output control transistor 214 in the off state.
.

~--~~r7rr.nrr~ r~

is, in turn, liill forc~ the ~)utput sta~es 216 throu~d on.
This asserts the output of the PO~J circuit lS2 in the zero state) the state indicating that power is not witllin acceptable limits.
.~s voltagc rises, the input voltage on line 190 will increase until it equals the reference voltage on linc lSS. At this point the differential amplifier lS6 drives the main output control transistor 214, turn llO ing it on. This removes the base drive from the output stages 216 through 228, forcing these output stages off. The output of the PON circuit 182 is then a one, indicating that the power is within acceptable limits.
At this point the hysteresis control circuit 230 comes into play. While power was coming on, the transistor 236 of the hysteresis control circuit 230 was on. l~hen the transistor 236 is on, the resistance value of the resistor 202 appears to be less than the resistance value of this resistor 202 is when the transistor 236 is off.
The point at which the main output control transistor 214 turns on is the point at which the ~` hysteresis transistor 236 turns off. Turning off the ` hysteresis transistor 23G causes a slight voltage jump ; in the line 190 which further latches the differential amplifier 186 into the condition wherc thc differential ~' ~

~36'7;Z8 amplifier 186 sustains the main output transistor 214 in the on state.
The state of the PON circuit will retnain stable in this condition with the main output control transis-tor 214 on and the output drivers 216 through 228 off until the plus five volts drops below a lower threshold point, as determined by the voltage applied on the line 190.
As the voltage on the line 190 decreases below the reference voltage on the line 188, (because the five volts supply is going down in a power failure condition), then the differential amplifier 186 turns off the main output control transistor 21~. This, in turn, turns on the output driver stages 216 through 228.
Since the hysteresis transistor 236 was off as power dropped, the voltage applied to the input of the PON circuit 182 must drop somewhat farther than the point at which the PON circuit 182 sensed that power was within the acceptable limits during the power~up phase of operation.
This differential or hysteresis is used to inhibit any noise on the five volt power supply from causing any oscillation in the circuit that would erroneously indicate that power is failing.
~he PON circuit 1~2 shown in Figure 25 provides - very accurate sensing of the two voltages used by the ':~

~ .

1~-1 PON circuit to determine its state (whether a on~ or a 2 ~ero output of the PON circuit).
4 In order to sense these two volta~es very 5 accurately the PON circuit must have the capability o 6 compensating for initial tolerancQs of the dif er~nt 7 components and also the capability to compensate for 8 changes in temperature durin~ operation. In the PO.J
9 circuit 182, the zener diode 192 is the only critical 10 part that must be compensated for because of its initial 11 tolerance, and this compensation is provided by selecting 12 the resistor 198.

14 Temperature compensation is achieved because 15 the zener diode 192 is an active zener diode and is 16 not a passive zener diode. Effective temperatur2 17 compensation is also achieved because the two transistors 18 in the differential amplifier 186 are a matched pair of 19 ~ransistors and the resistors 198, 200 and 202 are me~al 20 film resistors.

22 Each port 43 includes a number of lines which 23 are indicated by the general reference numeral 179 in 24 Fig. 20 and Fig. 19. This group of lines 179 includes 25 the individual lines 201 (sixteerl (16) of which make up 26 the Input Bus - I Bus), device address lines 203, Output 27 Bus lines 205 (of which there are sixteen), a take owner-28 ship line 207 and genéral llnes 209 which transmit such 29 signals as-parity, the T bus, and other similar lines 3~ which are required because o~ the particular hardware 31 implementation.

~L~L3~;7Z8 1 These Partlcular lines 201, 203, 205, 207 and ~ 209 correspond to the lines with the same numbers in 3 Fig. 21, which is the block diagram of the interface 4 common logic. ~owever, there are two sets of each of 5 these lines ln Fig. 21 ~ecause the interface common 6 logic 181 is associated with each of the dual ports 43 7 in a.device controller 41~

9 With reference to Fig. 21, the interface common 10 logic 181 includes the ownership latch 185 (see also 11 Fig. 19). This ownership latch determines the logical 12 connection between the interface common logic 181 and 13 a port 43 from which TAKE OWNE~SHIP signal has been 14 received over the line 207.

lS
16 As noted above, the TAKE OWNERSHIP signal is 17 derived by the port hardware from a load address and 18 command (LAC) T bus command (see Fig. 18) with a particular 19 operation code in the command field on the D bus. When 20 the port receives the function LAC on the T bus from the 21 channel, the port logic examines the command field (the 22 top six bits) on the D bus. Then, if the command field 23 contains a code specifying a take ownership command, the 24 port hardware issues a signal to set the ownership ~5 latch to connect the port to the interface common logic 26 and thence to the control part of the device controller.
27 If the command field speci~ies a kill command, the port 28 hardware issues a signal to clear the port's enable latch.
29 ~his operation happens only if the device addre~s field 30 on the D bus matches the port's device address jumpers, r ~1~3t;72~

1 and no parity error is detected during ~he command.
2 That i~, no commands (including the take ownership, 3 kill, etc.) are executed if a parity error is detected 4 on the LAC~

6 As a consequence, the I/O channel 109 7 issuing the Take Ownership command gains control of ~he 8 device controller 41, and the other port 43 is logically g disconnected. Take Ownership may also cause a hard 10 clear of the controller's internal state.

12 ~he state of the ownership latch 185 deter-13 mines which port may pass information through the multi-14 plexer 211. Once the ownership latch 185 is set in a 15 given direction, it stays in that state until a Take 16 Ownership command is received by the other port.

17 Assertion of the I/O reset line (IORST) will also cause 18 ownership to be given to the other port after the internal 19 state of the device controller has been cleared.

~1 _;

;7;~1 Control signals are chosen b~ the state of the ownership register 185 and from the appropriate one of the ports 43 and are transmitted by the mul-tiplcxer 211 to the control part 187 of a device con-troller on a set of control lines 215. Data are selected from an appropriate one of the ports ~3 on lines 205 and are loaded into the data register 213 and presented to the controller on an Output Bus ~O bus) 217.
Some of the control lines 215 (the lines 215A) are used to control the multiplexer 220 in selecting information from the controller as transmitted on lines 219, to be returned by the input bus (I bus) 201 to the ports 43 (Figure 20) and then to the channel 109 of a processor module 33. A line 221 returns the device address from the appropriate port 43 to the I bus 201 and thence to the I/O channel 109.
The data buffer 189 shown in Figure 19 is illustrated in more detail in Figure 22.
In accordance with the present invention many of the device controllers 41 incorporate a multi-word buffer for receiving information at a relatively slow rate from a peripheral device and then transmitting that information at or near memory speed to the processor module to maximize channel bandwidth utilization.
In the buffer design itself it i.s important that the device controllers ~]. be able to cooperate - 115 ~
B

..

f ~L~3f~7ZI~

1 with each othe~ in gaining access to the channel 109 ~ to avoid error conditions. In order for the de~Jice 3 controllers 41 to cooperate properly, the multiword 4 buffexs 189 are constructed to follow certain guideIines.
~ . , These guidelines include the followiny:

8 First of all, when a de~ice controller makes g a reconnect request for the channel 109 it must have 10 enough buffer depth left so that all higher priority 11 device controllers 41 and one lower priority device 1~ controller 41 may be ser~iced and the reconnect latency 13 of the reconnect request can occur without exhausting 14 the remaining depth of the buffer. This is called Buffer 15 Threshold, abbreviated T in Fig. 23.

~7 Secondly, after the buffer has been serviced, ~8 it must wait long enough to permit all lower priority 19 device controllers 41 to be serviced before making ZO another reconnect request. This is called Holdoff.
~1 ~he buffer depth (D in Fig. 23) is the sum of the holdoff 22 d~pth plus the threshold depth.

24 The holdoff and threshold depths are a 25 function of a number o~ variable~. The~e include the 2~ device rate, the channel rate, the memory speed, the 27 reconnect time, the number o controller~ of higher prioriky 28 0~ that I/O bus, the numbe~ of controller5 of lower priority 29 on that I/O bus, and the maximum burst length permissible.

., r 72~

1 A controller at high priority on an I/O bus has 2 more controllers of lower priority associated with it on ~he 3 same I/O bus than another controller at lower priority on 4 the same I/0 bus, and therefore the higher oriority controller 5 requires more holdoff depth than the lower priority controller.
6 Similarly, a controller at low priority on an I/O bus requires 7 more threshold depth than a controller at higher priority.
8 The buffer 189 in a controller is constructed to ~ake advantage g of the fact that as holdoff requirement increases the 10 threshold requirement decreases, and as the threshold 11 requirement increases the holdoff requirement decreases. This 12 is accomplished by making the stress at which a reconnect 13 request is made be ~aria~7e, the actual setting depending 14 on the characteristics of the controllers at higher and 15 lower priority in a particular I/O channel configuration.
16 The buffer depth is therefore the maximum of the worst~case 17 threshold depth or worst-case holdoff depth requirement, ~8 rather than khe sum of the worst-case threshold depth and 19 worst-case holdoff depth. This allows the buffer depth to 20 be minimized, and shortens the time required to fill or 21 empty the buffer.

2~

` 116a _, r ~L~3~i~2~

1 A number of these parameters are gr~phically 2 illustrated in Fig. 23. In Fig. 23 time has been plotted 3 on the horizontal axis versus words in the buffer on 4 the vertical axis for ~n output operation.

6 Starting at point D on the upper left hand part of Fig. 23 (and assuming a buffer filled to the full 8 buffer depth), data is transferred to a device at a rate 9 indicated by the line of slope -RD and this data transfer 10 continues without any reconnect signal being generated ~1 until the buffer depth decreases to the threshold depth 12 as indicated by the intersection of the line of slope -RD

13 with the threshold depth line ~ at point 223.

lS At this point the reconnect request is made 7~ to the channel 109 as indicated b~y the legend on the ., .
17 horizontal axis in Fig. 23.

19 ~he transfer of data continues from the bu~fer 20 at the rate indicated by the line of slope -RD and the 21 request is held off by higher priority device controllers 22 41 until point 225 at which point the request is honored 23 by the channel 109, and the I/0 channel begins its 24 ~econnect sequence or this device controller.

26 At point 227 the ~irst data word has been trans-27 mitted by the channel 109 to the device controller bufer ~8 189, and the channel 109 then transfers data words at 29 a rate indicated by the line of slope RC into the bufer 30 189.

. rf 113~7ZB

1 At the same time the device controller 41 2 continues to transfer data words out o~ th~ ~uf~er at 3 the rate -RD ~o that the overall rate of input to the 4 buffer 189 is indicated by the line of slope RC ~ RD
5 until thé buffer is again filled at the point 229. At 229 the buffer is ~ull, and the d~vice controller dis-7 connects from the channel 109, and the data transfer 8 continues at the rate indicated by the slope line -Rc.

The notation tr in Fig. 23 indicates the time 11 required for the polling and selection of this device 12 controller and the transfer of the first word~ This will 13 be discussed again below in relation to Fig. 16.

The letter B in Fig. 23 indicates the burst 16 time. The ~urst time is a dynamic parameter. The length 17 of any particular burst is dependent upon the device 18 transfer rate, the channel transfer rate, the number of 19 devices with transfers in progress and the channel 20 reconne~t time. The maximum tLme permitted for a burst 21 is chosen to minimize the amount of buffer depth required 22 while accomodating high device transfer rates and also 23 the number of devices that can transfer concurrently.

~5 Fig. 22 is a block diagram o~ a particular 26 embodiment of a buffer 189 constructed in accordance 27 with the present invention to accomplish the holdof~

28 and threshold requirements illustrated in Fig. 23.
~9 ~L3~Z~3 1 The buffer 189 sho~n in Fig~ 22 comprises an 2 input buIfer ~31, a buffer memory 233, an outpu~ bufer 3 235, an input pointer 237, an output pointer 239, a 4 multiplexer 241, ~uffer control logic 243 ~described in S more detail in Fig. 26), a multiplexer 245 connected to ; 6 the buffer control logic 243 and a stress counter 247.

8 As also illustrated.in Fig. 22, two yroups of g data input lines (lines 217 and 249) are fed into the 10 inpu~ buffer 231, : 12 One group of data input lines include sixteen 13 device data input lines 249.

The other group of input lines includP sixteen 16 Output Bus lines t bus lines) 217.

18 One or the other of these two groups of input g signals is then fed from the input buffer 231 to the 20 buffer ~emory 233 by a group of li~es 251. There are 21 sixteen of the lines 251.

Data is taken from the bufer memory 233 and 24 put into the output buffer 235 by a group o~ lines 253.
25 There are sixteen of the lines 253.

27 The output buffer 335 transmits the data 28 back to the interface co~non logic 181 ~see Fig. 19 29 and Fig. 21) on a group of sixteen li~es 219 and to 30 the devices 45, 47 (such as 49, 51, 53 shown in Fig. 1) '7Z8 1 on a group of sixteen lines 255 as indicated ~y the 2 legends in Fig. 22.
4 The input and output pointers 237 and 239 5 function with ~h~ multiplexer 241 as ~ollows.
7 When data is being kransferred from the 8 input buffer 231 to the buffer memory 233, the input 9 pointer 237 i5 connected to the buffer memory 233 10 through the multiplexer 241 to determine the location into which the word is written.

13 When data is being transferred out of the 14 buffer memory 233 into the output buffer 235, the output 15 pointer 239 is connected to the buffer memory 233 through 16 the multiplexer 241 to determine the location from which 17 the word is taken.

72~il 1 The purpose of th~ buffer control loyic 24 2 illustrated in Fig. 22 and Fig. 26 is to keep track o~
3 the stress placed on the buffer 189. In this reyard, 4 the degree of the full or empt~ condition of the buffer S in combination with the direction of the transfer with 6 respect to the processor module (whether lnput or output) 7 determines the degree of stress. Stress increases as the 8 device accesses the buf fer and decreases as the channel 9 accesses the buffer.

11 In the implementation shown in Figs. 22 and 26 12 the stress counter measures increasing stress from 0-15 13 on an input, and decreasing stress from 0-15 on an out-14 put. Another implementation ~not shown in the drawings) 15 would add the direction of transfer in the buffer control 15 logic such that two new lines would access the pointers `17 237 and 239 and the stress counter would always measure 18 increasing stress.

With continued reference to Fig. 22, a channel 21 request line 215 (see also Fig. 21) and a device request ~3 ~4 ~5 2g 3~

_ "

r ii7~8 1 line ~7 tcoming ~rom the control part lB7 of th~
2 device controller) are asserted ~o indicat~ acc~ss to 3 ~he buffer 189.
The multiplexer 245 chooses one of these lines 6 as a request to increase the buffer fullness and chooses 7 the other line as a request to decrease the buffer full-8 ness based on the direction of the transfer ~whether 9 input or output) with respect to the processor module.

11 The line chosen to increase buffer fullness 12 is also used to load data from the appropriate data 13 lines 249 or 217 (see Fig~ 22) into the input buffer 14 231 by means of the line 259.

16 The channel and the de~ice may ac¢ess the 17 buffer 189 at the same time, and the bufer control 18 logic 2'43 ser~ices one request at a time. The buffer 19 control logic 243 chooses one of the lines ~or service 20 and holds the other line off until the buffer control 21 logic 243 has serviced the first request, then it 22 services the other request.

24 The servicing of a request by the buffer 25 control logic 243 includes the following.

27 First of all, it determines the direction o~
28 transfer ~into or ouk of) the buffer memory 233, and it 29 asserts line 261 ~connected to the multiplexer 2~1) as 30 appropriate to select the input pointer 237 or the output 31 pointer 239 through the multiplexer 241.

~;367Z~

1 Secondly, on an output request, thc buff~r control 2 logic 2~3 asserts line 263 which does three thing.
3 (A) It writes the word from the input buffer Z31 4 into the buffer memor~ 233 at the location determined by 5 the input polnter 237 and the multiplexer 241.
6 (B) It increments the stress counter 247.
7 tC) The buffer control logic 243 increments the ~ input pointer 237.

Thirdly, on an output transfer, the buffer control lt logic 243 asserts line 265 which accomplishes the following 12 three operations~
13 (A) The buffer con~rol logic 243 writes the word 14 being read from the buffer memory 233 as determined by the output 15 pointer 239 and multiplexer 241 into the output buffer 235.
16 tB) The buffer control logic 243 decrements the 17 stress counter 247.
18 (C) The buffer control logic 243 increments the 19 output pointer 239.

21 The stress counter 247 determines when the buffer 22 `189 is full (D), or at threshold depth (T) as shown by the 23 output line legends in Fig. 22.

The output of the stress counter is decoded, and any 2~ one of the decoded values may be used to specify that the buffer 27 is at threshold depth. In the preferred embodiment, wire jumpers 28 are used to select one of sixteen possible stress values, and 29 a reconnect request is made to the channel lO9 when the stress 30 on the buffer 189 reaches that value.

1~3 r ~3~'72~

1 The control part 187 of the device con~roller 2 uses these three signals ~which correspond to the ley~nds 3 in Fig. 23) to mak~ reconnect requests and disconnect 4 requests on respective lines 145 (see Fig. 14 and Fig. 12) 5 and 159 (see Fig. 14 and Fig. 12).

7 The STI (stop in) signal transmitted on line 159 shown 8 in Fig. 14 and Fig. 12 is related to the buffer depth (D), the g full or empty conditions of the buffer and the direction of transfer;
10 and the RCI (reconnect in~ signal on line 145 of Fig. 14 and Fig. 12 11 is related to the threshold depth (~) indication from the stress ~2 counter 247 in Fig. 22. Thus, the STI signal is asserted when 13 the buffer 189 reaches a condition of minimum stress (full on output 14 an~ empty on input). The STI signal signals the channel 109 that 15 the controller 41 wishes to terminate the burst data transfer.
16 When the buffer passes through its threshold, it asserts the RCI
17 signal on line 145 to indicate to the channel 109 that the buffer 1~ wishes ~o transfer a burst of data.

Fig. 26 shows details of the multiplexer 245, the ~1 bu~fer control logic 243 and the stress counter 247 of the ~2 ~uffer 189 shown in Fig. 22.

24 In Fig. 26 the multiplexer 245 is shown as two sets of 25 gates 245A and 245B, request flip-flops 267A and 267~, a clock 26 flip-flop 269, request synchroniæation flip-flops 271~ and 271B, a 27 priority resolving gate 273 and request execution gakes 275A and 275E.

29 The stress counter 247 comprises a counter section 30 247A and a decoder section 247B as indicated by the legends 31 in Fig. 26.

1~4 3S~2~

1 As illustrated in Fiy. 26, the two sets o~
2 gates 245A and 245B have used the ch~nnel request siynal 3 (line 215~ and the device request signal (line 257) 4 and the read and write signals to determine which of the channel or the device is putting data onto the buffer 6 189 and which is taking data out of the ~uffer 189.

8 The xequest flip-flops 267A and 267B store the 9 requests until the con~rol logic has serviced the request.

11 The cloc~ flip-flop 269 generates a two phase 12 clock used by the request synchronization flip-flops 271A
13 and 271B and the request execution gates 275A and 275B.

The request synchronization flip-flops 271A
16 and 271B synchronize the request to the clock generation 17 flip-flop 269 and stabilize the request for execution.

19 The priority resolving gate 273 picks one 20 of the requests for execution and causes the other ~1 request to be held off.

23 The request execution gatés 275A and 275B
24 execute the requests in dependence on the synchronized 25 requesk.

27 Each output siynal on the lines 263 and 265 28 performs the functions described above (incrementing 29 and decrementing the stress counter, updating the 30 bufrer memory or output bu~fer, and updating the input 31 pointer or output pointer).

~36 ~

1 In addition, each signal clears the appropriat~
2 request flip-flop throuyh the lines 277A and 277B
3 illustrated in Fig. 26.
As noted above, Figs. 15, 16 and 17 show 6 the three sequences of operation of the I/O s~stem.
8 In the operation of the I/O system, the normal 9 data transfer between a processox module 33 and a particular device, such as a disc 45, includes an EIO
11 sequence to initiate the transfer.

13 The EIO instruction selects the particular 14 device controller and device and specifies the operation to be performed.

17 The device controller 41 initiates the I/O
18 between the device controller 41 and the particular 19 device.
2~
21 The device controller 41 periodically 22 reconnects tQ the channel 109 and transfers data 23 between the device controller 41 and the channel 109.
~4 The periodic reconnection may be or the purpose of either transferring data from the channel to the device 26 or for the purpose of transferring data from the device 27 to the channel.

29 When the transfer of data is compléte the device controller ~1 interrupts the CPU 105, which 31 responds by issuing an IIO or an HIIO sequence.

1~3~:i72~

1 The IIO sequence determines the idcntit~ of 2 the interrupting device and conditions under ~/hich th~
3 transfer completed.
Th~ HIIO sequence is similar to the IIO
6 sequence but is issued in response to a high priority 7 I/O interrupt.

9 The "Execute I/O" CPU instruction (EIO) is 10 defined by the T bus state changes sho~m in Fig. 15.
~1 12 The first state shown in Fig. 15 ~the state 13 farthest to the left) is the no-operation (NOP) or 14 idle stateO The other states are tXe same as those 15 listed in Fig. 18 by the corresponding mnemonics--load 16 address and command (LAC), load parameter (LPRM), read 17 device status (RDST), deselect (DSEL) and abort 18 instruction (ABTI).

As in the state changes shown in Figs. 6, 7 21 and ~, the solid line arrows indicate a state change, `22 and a dashed line arrow indicates a condition which must 23 occur before a state change can occur.

The EIO instruction and execution shown in 26 ~ig. 15 is directly under control o~ the microprocessor 27 113 (see Fig. 12) of the CPU 105.

29 This CPU initiation i~ shown as transmitted 30 to the state machine in Fig. 15 by the line 117; the ~3~

~ initiation siynal is accepted only when the T hus is 2 in -the idle s~ate.

4 Once the CPU initiatiGn signal is applied, the T bus goes from the NOP (idle) state to the LAC
6 state.

8 In the LAC state or function a word is taken g from the top of the register stack 112 in the CPU 105 tsee Fig. 12) and ls put on the D bus 161 (see Fig. 14~.
11 -, -. .
12 As described above, this word is used to 13 select a particular device controller 41 and a particular 14 peripheral device 45, 47, 49, 51 or 53 (see Fig. 1)~
and the word is also used to specify the operation to 16 be performed.

18 In the next T bus cycle the T bus goes to 19 the LPRM state.

21 In the load parameter state ~LPR~l) the word 22 just below the top of the register stack in the CPU
23 105 (see Fig. 12) is put on the T bus 161 (see Fig.
24 14) by the I/O channel 109 and i~ pa3sed to the device controller 41 selected during the previous LAC state.
2~
27 At the conclusion o the handsh~ke cycle, 28 as shown by the dashed line arrow in Fig. 15, the T
29 bus goes to the RDST state. In this state the device controller 41 returns the device status ~the status of 31 ~L36~Zl~

1 a particular devic~ selected and comprisiny the set 2 of signals describing the state of that device) from 3 the device controller 41 and places it on the top of 4 the register stack 112 in the CPU 105.
~ During the load parameter and read device 7 status state several errors may have occurred. These 8 include parity error, handshake time out, and an error 9 indication in the status word. If an error did occur, 1~ then the T bus machinP 143 (Fig. 13) goes from the 11 RDST state to the abort instruction (ABTI) state.
1~2 ~3 The ABTI state instructs the device controller 14 41 to ignore the previous LAC and LPRM information passed to it by the I/O channel 109 and then the T bus 16 (channel) returns to the NOP (idle) state.

18 If, after the RDST state no error was detected, 19 ~as shown by the dashed line arrow 114 in the top branch of Fig. 15), the T bus goes to the deselect state (DSEL).

22 With the ~ bus in the deselect state, the device 23 controller 41 clears its select latch 173 and responds to 24 the instruction issued to it (passed to it during the LAC
2~ state) and the T bus returns to the NOP (idle) state.

27 In the operation o~ the I/O system there are 28 a number of device request signals that can happen 29 asynchronously. For example, a reconnect signal may be 30 generated after an EIO sequence to request that the ~ ~ f .
1~l3S7;;~
1 channel transfer data to the controller. Or thc ~cvl~
2 controller 41 may assert an interrupt requ~st lin~ ~n(!~ir 3 a number of different conditions, e.~ to si(3~al th~
4 completion of an EIO sequence or to report ~n unusu~l condition in a peripheral device.

7 The device request lines are common to ~11 8 device controller ports 43 attached to ~ ~articul.lr I/O
9 bus 39.

11 ~he channel 109 responds to reconnect rcquests 12 made on the line RCI (145 of Fig. 14), and the CPU 105 13 responds to requests made on the LIRQ line 147 (s~e also 14 Fig. 14) with an IIO sequence, and to a request made on the HIRQ line 149 with an HIIO sequence.

17 The first thing that the channel 109 or CPU
18 105 doe5 in response to a Device Request signal is to 19 determine the identity of the highest priority device controller 41 asserting a requestO That is, there may 21 be several device controllers 41 asserting a request 22 to the channel 109 at one time, and the channel will 23 select a particular device controller in accordance 24 with a predetermined priority scheme.

26 In a particular embodiment of the presen~
27 invention up to thirty-two device controllers 41 can 28 be connected to a single channel 109.

1 The thirty-two device controllers are 2 connected in a star poll using the sixte~n bit data 3 bus 161. One additional line 151 is used to divide 4 the thirty-two device controllers into two groups of sixteen each. One group of sixteen devicc controll~rs 6 is assigned priority over the other group; and priority 7 is also assigned among the si~teen within each group.
8 The device responding on bit zero of the D bus during a 9 polling sequence has the highest priority within a rank, and the one responding on bit 15 has the lowest 11 priority.

13 In initial introduction, it may be noted that 14 polling (which will now be described) involves the state descriptions shown in Fig. 16 and 17 up to and including 16 that handshake which occurs during the select (SEL) state 17 in each figure.

19 With continued general reference to Figs. 16 and 17, the channel 109 sets the rank line to zero and 21 then presents the T ~us function RPOL (Fig. 16) if the 22 response is to a reconnect request, while the CPU 105 23 presents an LPOL (Fig. 17) T bus function if the CPU is ~4 responding with an IIO sequence, or an HPOL T bus ~unction if the CPU is responding with an HIIO sequence. This 26 is the only major polnt of difference be~ween the show-27 ings in Fig. 16 (the channel xesponse) and Fiy. 17 ~the 28 CPU response) with regard to polling.

!

~L~3~;7Z8 Referring specifically to Figure 16 ~n~ the response of the channel 109 to assertion of the RCI
line 145 (see Figure 1~), all devices with a reconnec-t reguest pending that would respond on rank zero place a one bit response on the D bus. That is, all these devices assert a llne of the D bus 161 correspondlng to thelr priority within the rank.
The channel 109 transfers the D bus response into the priority resolve register 135 (see Figure 13).
This priority resolve register 135 output determines wh.ich device controller has the highest priority (in accordance with the scheme described above) and asserts the appropriate bit back onto the D bus 161, if there is a bit asserted in rank zero by the attached device controllers.
If there are one or more devices asserting a response to the prlorlty resolve register on rank zero, the output of the priority resolve register is presented to all device controllers attached, along with the select function (SEL) on the T bus, and the device controller whose priority on rank zero matches the output of the priority resolve register sets its select bit 173 (see Figure 19), and then that port will respond to subsequent states in the sequence. Thi,s is the rnode of operation indicated by the solid line arrow going ~rom the state indicated by RPOL with a rank equals zero ko select (SEL).

~s'~
.., ~, "

:~l3~

1 If the priorit~ resolving register 135 2 determines that no device responded when the rank line 3 equalled zero, then the channel 109 sets the rank line 4 to one and reissues the RPOL T bus command. Then, if the priority resolving register determines that a 6 response occurr~d on ranX 1, the channel asserts the T
7 bus select function as be~ore.

g However, if the priority resolving register 1~ 13S deter~ines that no response was made on rank l, 11 the channel returns to the idle state indicated by 12 state NOP in Fig. 16.

14 This latter event is an example of a failure which might occur in one port 43 and which would result 16 in the system 31 accessing that particulax device 17 controller 41 through the other port 43.

g As noted above, the action of the priority resolving register 35 in response to an IIO or an HIIO
21 sequence initiated by the CPU 105 is the same as the 22 response of the priority resolving register 135 to a 23 reconnect sequence initiated by the channel in response 24 to a reconnect in on the line 145 ~rom a device controller 41.

27 With continued reference to Fig. 16, the 28 reconnect se~uence begins with the poll sequence described ~9 above for reconnecting the highest priority device controller 41 making a request.

;7Z~

1 The next st~p in the reconnect secIu~nce is ~ to determine the actual device controller number contained ; 3 in the device address comparator 133. As noted a~ov~, 4 the device address comparator 193 includes jumpers to determine a physical device controller number. These 6 are the same jumpers that are used on a LAC T bus 7 function during an EIO sequence to determine a particular 8 port. In the reconnect sequence the address determined 9 by these jumpers is returned to the I/O channel via the D bus during the T bus RAC state to access a table de-11 fining the buffer area for this device.

13 It is also necessary to determine the direction 14 Of the transfer (whether an input or outpu~ transfer to the processor module). To accomplish this determination 16 of the direction of the re~uested transfer and the device 17 address, the channel asserts the RAC T bus function and 18 the device controller 41 returns khe device controller 19 address and the transfer direction.
21 The channel uses the device address returned 22 by the device controller 41 to access a two word entry 23 ~142) in ~n I/O control table (IOC) 140 (Fig. 12) which 24 defines a buffer area 138 in the memory 107 or this particular device controller and device.

27 The format of a two word entry 1~2 is shown 28 enlarged in Fig. 12 to show details of the fields of the 29 two word S7.

1~36'7~8 There is a two word entry 142 in the IOC
table 140 for each of the eight possible devices of each of the thirty-two possible device controllers 41 attached to an I/O bus 39 associated with a particular processor module 33, and each processor module 33 has its own IOC table.
Each two word entry describes -the buffer location in main memory and remaining length to be transferred at any particular time for a particular data transfer to a particular device. Thus, as indicated by the legends in Figure 12, the upper word specifies the transfer address to or from which the transfer will be made by a burst; and the lower word specifies the byte count specifying the remaining length of the buffer area and the status of the transfer.
The fields representing the status of the transfer include a protect bit P and a channel error field CH ERR. The channel error field comprises three bits which can be set to indicate any one of up to seven numbered errors.
The transfer address and byte count are up-dated in the IOC table 140 at the conclusion of each reconnect and data transfer sequence (burst). The transfer address is counted up and the byte count is counted down at the conclusion of each burst. The amount reflects the numbex of bytes transferred during the burst, -X~

13~3~7~8 The seccnd word also con-tains (1) a fi~ld in 2 which any error encountered durin~ a r~conn~ct and dat~
3 transfer sequence ma~ be posted for later analysic, and 4 (2) a protect bit to specify that the buffer area in 5 memory 107 may be read from but not written into.

7 The protect bit serves to protect the processor 8 memory 107 from a failure in the device controller 41.
g That is, when the device controll~r 41 returned the lO transfer direction to the channel 109 during a read 11 address and command (RAC) T bus unction, a failure in 12 the device controller 41 could cause the devic2 controller 13 to erroneously specify an input transfer. Then the 1~ channel would go to the IN state and transfer data from ~5 the device controller into memory, thus causing data in 16 the buffer 138 to be lost. The protect bit allows the 17 program to specify that the channel may not write into 18 this bu~fer area; that is, the device may only specify 1~ an output transfer.

21 The transfer address specifies the logical 22 path 139B (see Fig. 12~.

~4 The channel places the tran3fer address in 2~ the channel memory address register 129 (see Fig. 13) 26 and places the ~yte count in the character count register 27 131 (see Fig. 13).

~L3~7~

1 Depending upon the direction of the trarlsfer, 2 (which the channel retrieved from the d~vice durin~ th~
3 RAC state shown in ~ig. 16~, the channel puts the T bus 4 in either the IN state or OUT state and tra~sfers d~ta 5 between the device controller 41 and memor~ 107 usiny 6 the channel memory address register 129 to speci~y the 7 logical path 139C (see Fig. 12). The channel memory 8 address register 129 and character count register 131 9 are updated with each word transferred during the burst 10 to reflect the next address in the buffer and the number 11 Of chaxacters yet to be transferred. At the conclusion 12 of a burst the contents of the channel memory address 13 register 129 and of the character count reglster 131 are 14 written into the IOC table 140.
16 In operation, for each word transferred in 17 from the device on an in transfer, the channel 109 18 accepts,the word by the handshake mechanism described 19 above and places the word in the I/O data register 127 20 t9ee Fig. 13) and then transfers the word to the buffer 21 area in memory defined by the lo~ical path 139C (see 22 Fig, 12).

24 . On an out transfer the channel 109 takes a a 25 word from the buffer area over logical path 139C and 26 transfers the word to the channel memory data register 27 125~ The channel then transfers the wora into the I/O .
28data register 127 (Fig. 13) and hand.~hakes with the device 29controller which accepts the word into its interface .
30data register 213.

.:
~ 137 , ~L~3~'7Z~
1 The high sp~ed of the I/O chann~l is 2 accomplished by pipelining ~here the word in the I/O
3 data register 127 is handshaken to the device while the 4 channel concurrently requests and acc~ts the n~xt ; 5 word in the transfer from memory 107 and places it in 6 the channel memory data register 125. Since it ta~es 7 just as long to put a word out to the device as it does 8 to accept a word from memory for the device, the two g operations can be overlappea.

11 During the burst, the channel decremented the 1~ character count register by two for every word transferred, 13 since there are two by~es in every word.

The burst transfer can terminate in two ways.
`1~ The bùrst transfer can terminate normally or 17 the burst transfer can terminate with an error condition.

19 In the normal case there are two possibilities.

21 ~n a first condition of operation, the 22 character count register 131 can reach a count of either 23 one or two bytes remaining to be transferred. In this 24 situation the channel puts up EOT (line 165 as shown 2~ in Fig~ 14) signifying that the end o transer has 26 been reached. If the count reaches one, then the channel 27 asserts EOT and PAD OUT ~line 167 of Fig. 14) signifying 28 the end of transer with an odd byte.

If thc charac~cl cuurlt rca~les t~JoJ ~hc ch.lllnol puts Up EOI, but PAD OUT (I'/~DO on linc IG7 o~ Figllrc 14~
is not ~c~luired because both bytcs on the bus are valid.
In cither case, the devicc controllcr 41 responcls by asserting STOP IN (STI) on line 159 ~see Figurc l4) J and the device controller 41 also asserts P.~D I~' (PADI) on line 169 (Pigure 14~ i.f the channel asscrted PAD W T (PADO~.
In this first case of normal termination, the trallsfer as a ~hole, not just the burst, is terminated by tl~c channel 109.
The other normal completion is when the device controller 41 ends the burst by asser~ing STOP
IN (STI) in response to the channel SERVICE W T (SVO~.
This signifies that the buffer 189 (see Figure 19~ has reached a condition of minimum stress ~as indicated by point 22~ in Figure 23).
The STOP IN ~STI) can occur on an output transfer or on an input transfer.
~0 On an input transfer, if the device controller 41 wishes to terminate the transfer as well as the burst, the device controller 41 can assert STOP IN
~STI); and, to signify an odd byte on the last word, ~he device coDtroller 41 can also assert PAD IN (PADI).

~. ' ' `
.

r. 1' ,, 67~8 1 ¦ As shown in Fig. 16, when th~ tr~nser is 2 tl~rminated by a non-error condition (STI OR EOT) on 3 ~ither an output transfer or an input transfer ~as shown 4 ~y the balloons OUT and IN in Fig. 16), the channel l109 updates the IOC table entries as noted above, and 6 ¦returns to the idle (NOP) state shown in Fig. 16.

81 As noted above, the transfer can also be terminated by an error condition.

During the burst several errors may occur as follows.

4 First, the device controller 41 may request an input transfer into a buffer whose protect bit P is 16 set in the IOC table as mentioned above.

18 Second, the device controller 41 may not 19 return a PAD IN tpADI) signal in response to a PAD OUT
tpADo) signal from,the channel 109.

22 ~hird, the channel 109 may detect a parity 23 error on the D bus 161.

Fourth, the device controller 41 may not 26 respond to a SERVICE OUT ~SVO) signal from the channel 27 109 within the allotted time a~ mentioned above in the 28 discussion on handshakes.

, _,,.

.

1 Fifth, the buffer area specified b~ the IOC
2 table entries may cross into a page w~ose map marks it 3 absent (see the discussion of the mapping scheme in 4 the memory system).

6 Sixth, a parity error may be detected in 7 accessing the map while accessing the memory during 8 the reconnect in and data transfer sequence. See the 9 description in the memory system relating to the parity 10 error check.

12 Seventh, the memory system may deteci an un-13 correctable parity error when the channel 109 accesses 14 the memory. See the description of the memory system for 15 this parity error check.

`17 I~ any of these error conditions occur, 18 the channel 10g goes to the abort data transfer state 19 tABTD) as shown in Fig. 16. This instructs the device 20 controller 41 that an error has occurred and that the 21 data transfer should be aborted. The channel 109 then 22 goes back to the idle state which is (NOP) as shown in 23 Fig. 16.

When an error occurs, the channel 109 updates 26 the IOC table entries and puts an error number indicating 27 one of the seven errors noted above in the error ~ield 28 Of the second word of the IOC table entry as mentioned 29 above.

' ~3~c;72~

1 Thus, if a single error occurs, the ~umbcr of 2 that error is entered in the error field of the IOC
3 table entry.

If more than one error occurs, the chann~l 6 109 selects the error from which recovery is least 7 likely to occur and enters only the number of that error 8 in the error field of the IOC table entry.

There is one other type of error that can 11 occur. The device controller 41 may try to reconnect 12 to the channel when the count word in the IOC table is 13 zero. In this event, the channel will not let the 1~ device controller reconnect and the channel goes through the sequence as described above with reference 16 to Fig. 16, but when the channel determines that the -17 count word in the IOC tab~e is zero, the channel 109 18 goes directly to the abort (ABTD) state. This is an 19 important feature of the present invention because it ~o protects the processor memory from being overwritten by 21 a failing device.

23 If the count is zero in the byte count count 24 Of the second word of the IOC table entry 142 for a 25 particular device, and if the devlce controller 41 26 attempts to reconnect to the channel 109, the channel 27 issues an abort (ABTD) to ~he device controller 41 as 2~ noted above and leaves the channel error ~ield o~ the 29 two word entry 142 at zero.

~r, f ~3~72~

1 In response to an abor~ data (AB~D) T bus 2 function, the device controller 41 makes an interrupt 3 request on the line HIRQ or LIRQ (lines 149 or 147 as 4 shown in Fig. 14) to the channel 109.

6 The device controllers 41 may at any time 7 request an interrupt on these two llnes.

9 An interrupt generally indicates that a 10 data transfer has been completed or texminated by an 11 abort from the channe~ (an ABTD from the channel) or 12 by an error co~dition wlthin the device controller 41 13 or attached device, or that a special condition has 14 occurred within the device controller or an attached 15 de~ice. For example, when the power is applied and the 16 PON circuit indicates that power is at an acc~ptable 17 level, the device controller interrupts the processor 18 module ~o indicate that its internal state is Reset 1~ because power was off or had failed and has been reset 20 by the PON circuit.

22 In response to an interrupt, the program 23 running within the processor module 33 issues an interrogate 24 I/O instruction (IIO) or an in~erroyate high priori~y 25 I/O instruction (HIIO) over the I/O bus 3g.

27 The IIO instruction is issued in response to 28 a low priority I/O interrupt, that is, one issued on 29 the low priority interrupt request (LIRQ) line 147 (see 30 Fig, 14).

'.~3~

1 The ~IIIO instruction ls issued in response 2 to a high priority I/O interrupt, th~ is, one requested 3 on a high priority interrupt request (~7IRQ) line 149 4 (see Fig. 14).

~ The microprocessor 113 (see Fig. 1~) executes 7 the EIO, IIO or HIIO instruction by t~king control of 8 the channel control logic 141 and data path logic 123.

The sequence or these instructions is 1~ illustrated in Fig. 17; and, as noted above, the sequence 1~ starts with a polling sequence.

14 The IIO instruction polls in a sequence using the T bus function low priority interxupt poll (LPOL) 16 while the HIIO instruction polls in a sequence using 17 the T bus function high priority interrupt poll tHPOL).

19 The polling sequence which is also described above completes by selecting the appropriate device 21 controller 41 by using the T bus function select (SEL) 22 as shown in Fig. 17.

24 The appropriate device controller 41 selected is that device controller which has the hiyhest priority 26 and is making an interrupt request.

2g ~he sequence continues with a read interrupt 29 cause ~RIC) T bus function as shown in Fig. 17. The 30 device controller 41 responds by returning device 31 dependent status on the D ~us 161 (see Fig. 14).

f ~r 1 The microprocessor 113 (Fig. 12) reads the 2 status from the D bus 161 and places the status on the 3 top of the register stack 112 (Fig. 12).
The sequ~nce then continues with a read 6 interrupt status ~RIST) T bus function as sho~m in 7 Fig. 17. The device controller 41 responds to this RIST
8 T bus function by returning the device controller number, 9 the unlt number and four dedicated status bits on the D bus.

12 Of the four bit status field, two of the bits 13 indicate respectively, abort (ABTD) and parity error 14 (which parity error may have occurred during a reconnect ~5 and data transfer sequence).
~6 17 The microprocessor 113 copies the content of 18 the D bus- the controller number, the device number 19 and the interrupt status--and places that content on the 20 top of the register stack 112.
21;
22 If no error occurred during the sequence, then 23 the sequence continues with the deselect ~DSEL) state 24 which deselects the device controller 41; and then the 25 sequence goes into the idle (NOP) state as lndicated by 26 the line at the top of Fig, 17.

28 If an error did occur (and the error can be 29 a parity error detected by the channel or a handsha~e 30 tLme out), the channel goes from the RIST state to the 145 _, abort instruction (!~BII) jtatc as ;llowr~ ig~lrc 17.
Tllis dcselects th~ ~levic~ contro11er 41, alld then thc channcl 109 ~oes b~ck into the i~le ~oP) state as sho~in by the bottom line in Figurc 17.
~ s noted above, an -L/0 operation between a processor module and an I/0 dcvice typically consists of a group of sequences, e.g. an EI0 followed by soTne number of reconnect and data transfer sequences, terminat-ing witll an II0 sequence. Sequences from several different I/0 operations may be interle~ved, resulting in apparent simultancous I/0 operation by several devices. Thus, a large number of devices may be accessed concurrently;
the exact number depends on the channel bandwidth and the actual bandwidth used by each device.
The I/0 system and dual port device controller architecture and operation described above provide a number of important benefits.
These benefits include (a~ flexibility to interface a wide variety of devices, (b) a maximum usage of resources, ~c) a fail soft environment in which to access peripheral devices in a multiprocessor system, (d) on line maintenance and upgrade of the multiprocessor system capability, and (e) maximum system through put (as opposed to emphasizing processor through put or I/0 through put exclusively~ in an on line transaction system in which a large number of concurrent transactions must be processed by the I/0 system and CPU.

1 Flexibility to interface a wide variet~ of 2 devices is achieved because the system of the pre~ent 3 invention does not presuppose any inherent characteristics 4 of a device type. Instead, the present invention provides 5 a structure and operation which can accon~odate a wid~
6 variety of device operations.
8 The present invention provides for a maximum 9 usage of resources, primarily by making a maximum usage 10 of memory bandwidth. Each device uses a minimum of the ll memory bandwidth. ~his allows a relatively large number 12 of devices to be associated with the particular I/O bus.
13 Because of the inherent speed of the I/O bus, and the 14 buffering technique of the present invention, each 15 particular transfer is made at a relatively high speed 16 limited only by memory speedO Because the transfers are 17 in a burst mode, the overhead associated with each trans-18 fer is minimized. This maximizes the use of the channel 19 bandwidth and also permits the use of high speed devices.
21 The present invention provides for failsoft 22 access to peripheral devices. There are redundant paths 23 to each peripheral device, and contalnment of failure 24 on any particular path. Failure of a particular module 25 in one path doe~ not affect the operation o~ a module 26 in another path to that device.

28 There are comprehensive error checks for 29 checking data integrity over a path, sequence ~ailures 30 and timing failures.

~1~7 6"~2~ ~

Protectio~l ~eaturc~ l)revcnt a ~eriphe~al devicc from ~ontamill~tirlg its o~n buf~er o-r the memor~ of the system.
These protection fcatures include a separate eount word in each IOC table and a protcct bit in the IOC table. The IOC
tablc is accessible by the channel, but not by the device.
This is a second levcl of protcction to prevent the dcvice from acc~ssing any memory not assigned to that device.
The present invention requires only a small nwnber o~ lines in the I/O bus to provide a flexible and powerful I/O system.
The operation of the device controller is well defined as power is turned on or off to protect the I/O bus from erroneous signals during this time and also to permit on line maintenance and system upgrade.
The present invention uses stress to allow the buffers to cooperate without communicating with each other.
An on line transaction system is obtained through overlapped transfers and processing.
Multichannel direct memory access provides interleaved bursts to give overlapped transfers and minimum waits for accesses to a device. Each burst requires a minimum memory overhead and allows the pro-cessor to make maximum use of the memory. This combination allows maximum use of the I/O bandwidth and minimal tie up of the procossor, - 1~8 -3~

1 POWER DIS~RIBUTION S~STr~

3 The multiprocessor system of the present 4 invention incorporates a power distribution system that 5 over comes a number of problems associated with prior 6 art systemS.

8 In many prior art systems it was necessary to g stop the processor system in order to perform required 10 maintenance on a component of the system. Also, in many 11 prior art systems, a failure in the power supply could 1~ stop the entire processor system.

14 The power distribution system of the present l~ invention incorporates a plurality of separate and 16 independent power supplies and distributes the power 17 from the power supplies to the processor modules and to 18 the device controllers in a way that permits on line g maintenance and also provides redundancy of power on 20 each device controller.
~1 1 In this regard "on-line" is used ih the sense 2 that when a part of the system is on line, th~t part of 3 the system is not only powered on, but it is also function~
4 ing with the system to perform useful work.

6 The term "on-line maintenance" therefore means 7 maintaining a part of the system (including periodic 8 preventative maintenance or repair work) while the 9 remainder of the system is on-line as defined above.

11 In the present invention any processor module 12 or device controller can be powered down so that on-line 13 maintenance can be performed in a power off condition 14 on that processor module or a device controller while the 15 rest of the multiprocessor system is on-line and functional.
`16 The on-line maintenance can be performed while ully 17 meeting Underwriters Laboratory safety requirement~.

~1 _ `

1 Also, in the power distrlbution system of 2 the present invention each device controller is co~nected 3 for supply of power from two separate power supplies 4 and by a diode switching arrangement that permits the 5 device controller to be supplied with power from both 6 power supplies when ~oth power supplies are operative and 7 to be supplied with power from either one of the power 8 supplies in the event the other power supply fails; and 9 the changeover in the event of failure of one of the lO power supplies is accomplished smoothly and without any ll interruption or pulsation in the power suppl~ so that an 1~ interrupt to a device controller is never required in 13 the event of a failure of one of its associated power 14 5upplies.

16 A power distribution system for insuring both 17 a primary supply and an alternate power supply for each 18 individual dual port device controller 41 is illustrated 19 in Fig. 30. The power distribution system is indicated 20 generally by the reference numeral 301 in Fig. 30.

22 The power distribution system 301 insures 23 that each dual port device controller 41 has both a 24 primary power supply and an alternate power supply.
25 Because each device controller does have two separate 26 and independent sources o~ power supply, a failure o~ !
27 the p~imary power supply for a particular device controller 28 does not render that device controller (and all o~ the 29 devices a~sociated with that controller) inoperative.
30 Instead, in the present invention, a switching arrange-3~

ment provides for an automatic ~witchover to the alternate power supply so that the device controller can continue in operation. The power distribution system thus coacts with the dual port system of the device controller to provide uninterrupted operation and access to the devices in the event of a failure of either a single port or a single power supply.
The power distribution system 301 shown in Figure 30 provides the further advantage that each processor module 33 and associated CPU 105 and memory 107 has a separate and independent power supply which is dedicated to that processor module. With this arrangement, a failure of any one power supply or a manual disconnection of any one power supply for repair or servicing of the power supply or associated pro-cessor module is therefore limited in effect to only one particular processor module and cannot affect the operation of any of the other processor modules in the multiprocessor system.
The power distribution system 301 shown in Figure 30 thus works in combination with the individual processor modules and the dual port device controllers to insure that a failure or disconnection of any one power supply does not shut down the overall system or make any oE the devices inefEective, The power distribution system 301 includes a plurality of separate and indepenclent power supplies ~,-- , ~

~3L3~21~

1 303, and each power supply 303 has a llne 305 (actuall~
2 a multiline bus 305 as shown in Fig. 33) which is 3 dedicated to supplying power to the CPU and memory of 4 a particular, related processor module.

6 Each device controller 41 is associated with 7 two of the power supplies 303 through a primary line 307 8 and an alternate line 309 and an automatic switch 311.

A manually operated switch 313 is also 11 associated wi~h each device controller 41 between the 12 device controller and the primary line 307 and the 13 alternate line 309.

The switches 311 and 313 are shown in more 16 detail in Fig. 31.

18 Fig. 32 shows details of the component 19 construction of a power supply 303.

21 As shown in Fig. 32, each power supply 303 ~2 has an input connector 315 for taking power from the --23 mains. The input 315 is connected to an AC to DC
24 converter 317, and the output of the AC to DC converter 25 provides, on a line 319, a five volt interruptable 26 power supply (IP5). This five volt interruptable power 27 supply is supplied to the CPU 105, the memory 107 and 2~ the device controllér 41. See al~o Fig. 33.

153 ~.

. ,. ~. . .. . ...

-:~3~

The AC to DC converter 317 also provides on a second output line 321 a sixty volt DC outpu-t which is supplied to a DC to DC converter 323. See FicJure 32, The DC to DC converter in turn provides a five volt output on ~ line 325 and a twelve volt output on a line 327.
The outputs from the lines 325 and 327 are, in the system of the present invention, uninterruptable power supply (UPS) outputs in that these power supply outputs are connected to the CPU and memory when semi-conductor memory is used. The power sùpply to a semi-conductor memory must not be interrupted because a loss of power to a semiconductor memory will cause loss of all data stored in the memory.
The five volt interruptable power supply on line 319 is considered an interruptable power supply because this power is supplied to parts of the multi-processing system in which an interruption of power can be accepted. Thus, the five volts interruptable power is supplied to parts of the CPU other than semi-conductor memory and to only those parts of the memory which are core memory (and for which a 10.5s of power does not cause a loss of memory) and to the device controller which (as will be described ;n more detail below) is supplied with an alkernate sourcc of power in the event of a failure of the primary power supply.

- 15~ -Since the power supply on lines 325 and 327 must be an uninterruptable power supply, the pre~ent invention provides a batter~ back-up for the input to th~ DC to DC converter 323. This battery back-up includes a battery and charger module 32g. The module 329 is connected to the DC to DC converter 323 by a line 331 and a diode 333.
In a particular embodiment of the present invention the battery 323 supplies power at 48 volts to the converter 323, which is within the input range of the converter 323.
The diode 333 insures that power from the battery is supplied to the converter 323 if the voltage on the line 321 drops below 48 volts. The diode 333 also stops the flow of current from the battery and the line 333 when the output of the AC to DC converter on line 321 exceeds 48 volts.
Each power supply 303 also includes a power warning circuitry 335 for detecting a condition in the AC power input on line 315 that would result in in-sufficient power out on the output lines 319, 325 and 327. The power warning circuit 335 transmits a power failure warning signal on a line 337 to the related CPU 105.
Because of the capacity stor~ge in the power suppl~ 303, there is enouyh time between the power warn-'~ , 1 ing signal and the loss of the ~ive ~olts interruptable 2 power on 1ine 319 ~or the CPU to s~ve its state 3 before the power is lost.
~owe~er, the uninterruptable power supply on 6 lines 325 and 327 must not be interrupted, even for 7 an instant of time; and the battery ~ack-up provided 8 by the arrangement shown in Fig. 32 insures that there ~ is no interruption in the power supply on lines 325 10 and 327 in the event of a power failure in the input 11 line 315.

13 One particular power supply 303 itself can 14 fail for some reason with the other power supplies 303 15 still operating. In that event, the power distribution 16 system 301 of the present invention limits the effect 17 of the failure of the power supply 303 to the loss of 18 one particular, associated CPU and memory; and the auto-19 matic switch 311 provides for an automatic switchover 20 from the failed power supply to the alternate power 21supply to keep the associated device controller 41 in 22operation. The device controller 41 which had been 23connected to the failed power supply therefore continues 24in operative association with the other processor modules 25and components of the multiprocessor system, because the 26re~uired power is automaticall~ switched in ~rom the 27alternate power supply.

29 As best illustrated in Fig. 31, each automatic 30Switch 311 includes two diodes--a diode 341 associated with 1 the primary power lin~ 307 and a diode 3~3 associated ~ith 2 the alternate power line 309.

4 The function of ~he diodes 341 and 343 is to per-5 mit power to be supplied to a device controller 41 from 6 either the primary power line 307 and a related power suppl~
7 303 or the alternate power line and its related power supply 8 303 while keeping the supplies isolated~ This prevents a 9 failed power supply from causing its associated alternate 10 or primary from failing.
1} .....
12 In normal operation each diode permits a certain 13 amount of current to flow through the diode so that the 14 power to each device contrcller 41 is actually being 15 supplied by both the primary and alternate power supplies 16 ~or that device controller.

18 ~ In the e~ent that one of the power supplies lg fails, the full power is supplied by the other power supply, 20 and this transition occurs without any 10s5 of power at all~

22 Since there is a small voltage drop across the 23 diodes 341 and 343, the voltage on the lines 307 and 309 24 must be enough higher than five volts to accomodate the 25 voltage drop across the diodes 341 and 343 and still 26 supply exactly five volts to the device controller 41.
27 The lines 305 are in parallel with the lines 307 and 309, and the power actually received at the CPU in memory must 29 also be five volt5; so balancing dlode.s 339 are located 30 in the lines 305 to insure that the voltage after the diodes 31 339 as supplied to each CPU is exactly five volts.

The manual switch 313 permits a device controller 41 to be disconnected from both the primar~
and the alternate power sources when the device con-toller needs to be disconnected for removal and service.
Details of the construction of the switch 313 are shown in Figure 31. As shown in Figure 31, the switch 313 includes a manual switch 345, a transistor 347, a capacitor 348 and a resistor 350 and a resistor 352.
The manual switch 345 is closed to turn on the transistor 347 which then supplies power to the device controller 41.
. It is important that both the turn on and the turn off of power to the device controller 41 be accomplished in a smooth way and without fluctuations which could trigger the PON circuit 182 more than once.
The feedback capacitor 348 acts in conjunction with the resistor 352 to cause the required smooth ramp build-up of power when the switch 345 is closed to turn the trans-istor 347 on.
When the transistor 347 is turned off by open-ing the switch 345, the feedback capacitor 348 acts in conjunction with resistor 3~0 to provide a smooth fall off of power.

[n a l~rcfcrrcd elnbodilllcnt o~ thc irtvcntion all of diocles ~l, 343, and 3.,'J are Schottky diodes ~hich have a very low fon~ard volta~c drop7 and this rc~uccs power dissipation, As noted above in the description of the I/0 systcm and dual port devicc controller 41, each devicc contro11er 41 does have a powcr on circuit ~PON) lS ~or detecting when the five volt power is below spccifications. The P0~ circuit 182 is shown in more dct~il in Figure 25 and resets the device controller 41 to loc~ everything off of the device controller and holds the device controller itself in a state that is ~nown when the power is turned off by the switch 313.
The P0~ circuit 182 also releases the device controller and returns it to operation after the power is turned on by switch 313 and five volt power supply at the proper specification is supplied to the device controller 41.
Further details of the power on circuit 182 shown in Figure 25 are described above in relation to the I/0 and dual port controller system.
With reference to Pigure 33, the power from each power supply 303 is transmitted to a related CPU
by thc vertical bus 305, and each vertical bus 305 is a laminated bus bar which has five layers of electrical conductors.

As indicated by the leyends in Figure 33, each vertical bus 305 has two different conductors con-nected to ground.
One conductor provides the ground for both the five volt interruptable power supply (IPS) and the five volt uninterruptable power supply (UPS).
A separate conductor provides a ground for the memory voltage. This separate yround for the memory voltage insures that the relatively large fluctuations in current to the memory will not have any effect on either the five volt IPS or the ~ive volt UPS supplied to the CPU.
The horizontal bus 307, 309 includes the primary and alternate power supply lines 307 and 309 (as indicated by the reference numerals in Figure 30).
In a particular embodiment of the present invention the bus 307, 30g is actually a nine layer laminated bus which has a single ground and eight voltage layers (Vl through V8 as indicated by the legends and nota-tions in Figure 33).
Each voltage layer is connected to the five volt interruptable output of a different power supply 303. Thus, the layer Vl is connected at 351 to the five volt IPS power for the power supply 303 and related processor module farthest to the left as viewed in Fig-ure 33, and the layer V2 i8 connected at 353 to the five L~

"~

:~3~

volt IPS power supply 303 for the processor module at the center as viewed in Figure 33, and so on.
Since there are eight layers (Vl throuyh V8) and a common ground available to each device controller in the horizontal bus, upstanding vertical taps 355 to these eight layers at spaced intervals alony the horizontal bus permit each device controller 41 to be associated with any two of the power supplies 303 merely by connecting the primary line 307 and the alter-nate line 309 to a particular set of taps. By way of example, the device controller 41 on the lefthand side of Figure 33 is shown connected to the taps Vl and V3 and the device controller 41 on the righthand side of ~igure 33 is shown connected to the taps V2 and V3.
Thus, any device controller 41 can be con-nected to any two of the power supplies 303 with any one of the power supplies serving as the primary power supply and any one of the other power supplies serving as the alternate power supply.
The power distribution system of the present invention thus provides a number of important benefits.
The power distribution system permits on line maintenance to be performed because one processor module or device controller can be powered down while the rest of the multiprocessor system is on line and functional.

;3 1 The power distribution system full~ meets all 2 Vnderwriter Laboratory safety requirements fcr doing on 3 line maintenance or a powered down component while the 4 rest of the multiprocessor system is on line and in 5 operation.

7 Each device controller is associated with two 8 separate power supplies so that a failure in one of the g power supplies does not cause the device controller to 10 stop operation. Instead, the electronic switch arrange-11 ment o~ the prPsent in~ention provides such a smooth 12 transition of power from the two power supplies to only 13 one of the power supplies that the device controller is 14 maintained in conkinuous operation without an interrupt.
lS

~6 2g , 3 Each processor module 33 (se~ Fig. 1) in thc b~ multiprocessor system 31 contains a memory.
6 This memory is indicated by the ~en~ral 7 reference numeral 107 in Fig. 1 and is shown in greater 8 detail in Fig. 34.

The memory 10? of each processor module 33 11 is associated with both the CPU 105 and the I/O channel 12 109 of that module. There is a dual port access to 13 the memory by the CPU and the channel. That is, the 14 CPU 105 (see Fig. 1 and Fig. 34) can access the memory 15 for prograrn or data references, and the I/O channel 16 109 can also access the memory directly (without having 17 to go through the CPU) for data transfers to and from a 18 device controller 41. This dual access to the memory 19 is illustrated in Fig. 34 and will be described in 20 greater detail below in the description of the Fig. 34 21 structure and operation.

23 One benefit of this dual access to the 24 memory is that CPU and channel accesses to the memory 25 can be interleaved in time. There is no need for 2~ either the CPU or the charmel to wait ~or access to 27 the memory, except in the case where both the CPU
28 and the channel are tryiny to access the mernory at 2~ exactly the same time. As a result, both the CPU and 30 the channel can ~e performing their separate functions !

~3~8 1 simultaneously, subject to an occasional ~lait by the 2 CPU or channel if one of these units is accessing 3 the memory at the exact time the other unit needs to 4 access the memory.
6 The dual port access also allows background 7 I~o operation. The CPU 105 needs to be involved with 8 the channel 109 only in the initiation and termination g of I/O data transfers, The CPU can be performing other 10 functions during the actual I/0 data transfer itself.

12 The memory lb7 shown in FigO 34 comprises a ; 13 physical memory which consists of up to 262,144 words 1~ Of sixteen data bits each.

16 In addition to the sixteen data bits, each 17 word in memory has an additional parity bit if the 18 memory is a core memory or six additional error correction 19 bits if the memory is a semiconductor memory.
21 The parity bit permits detection of single 22 bit errors.

24 The six error correction bits permit detection 25 and correction o~ single bit errors and also permit 26 detection o~ all double bit errors.

28 The physical memor~ is conceptually subdivided - 29 into contiguous blocks of 1024 words each (which are 30 called pages~. The payes in physical memory are n bered , 164 ~L3~;7~

1 consecutively from page zero, starting at physical 2 location zero. The address range of physic~1 m~mory 3 in one specific emhodiment of the present invention, 4 which address range is zero through 262,143, requires eighteen bits of physical address information.

7 The basic architecture of the present 8 invention is, however, constructed to accommodate and g utilize twenty bits of physical address information, as 10 will become more apparent from the description to follow.

12 In one specific embodiment of the invention the physical memory is physically divided into physical 14 modules of 32,768 words. Thus, eight of these modules 15 provide the 262,143 words noted above.

17 All accesses to memory are made to one of 18 four logical address areas--user data, system data, user 19 code and system code areas. All C~U instructions deal 20 with these logical (as distinct from physical~ addresses 21 exclusively. Thus, a programmer need not be concerned 22 with an actual physical address but can instead write 23 a program based entirely on logical addresses and the 24 logical addresses are translated by the map section of 25 memory system into physical addresses.
~6 27 The range of addressing in any given logical 28 address area i~ that of a sixteen bi~ logical address, 29 zero through 65,535. Thus, each logical address area 30 comprises sixty-four logical pages of 1024 words each.

f r ~L~3~

1 In the memory system of the pre.sent 2 invention there i3 no required correspondence between 3 a logical page and a physical page. Instead, -the 4 various logical pages comprising an operating system 5 program or a user program need not reside in conticJuous 6 physical pages. In addition, the logical pages need 7 be in physical main memory but may be in secondary 8 memory, such as on a disc.

This allows implementation of a virtual 11 memory scheme.

13 Virtual memory has two benefits.

lS Fixst, ~irtual mem4ry allows the use of a physical 16 main memory space which is smaller than the logical address 17 areas would require, because the physical memory can be ~8 supplemented by a secondary physical memory.
lg Secondly, virtual memory permits address spaces 21 Of a pluralit.y of users (multiprogramming~ to share the 22 physical memory, and each user does not ha~e to be con~ ~~
23 ~erned with the allocation of physical memory among the 24 operating sys~em, himsel~, or other User~.

26 The memory system of the present invention 27 provides protection between users in the multiprogramming 28 environment by guaranteeing that one u~er progrnm cannot 29 read from or write into the memory space of another 30 user program. This is accomplished by the paging and .

~ ".

3~l3~

1 mapping system. When one ~ser program is runnin~, the 2 map fox that user program points only to the m~mory pages 3 (up to sixty-four pages of code and sixty four payes of 4 data) for that particulax user program. That particular 5 user program cannot address outside its own logical 6 address space and therefore cannot wri~e in~s or read 7 from the memory space of another user program.

9 The fact that code pages are non-modifiable lO also prevents a user program from des~roying itself.

12 Thus, there a~e two levels of protection for 13 user programs-operating in a multiprogramming environment--14 the fact that each user map points only to its own pages 15 in memory and the fact that code pages are non-modifiable.
16 Also, in the present invention, this protection is 17 achieved without prstaction limit registers or by pro~ection 18 keys as'often used in the prior art.

~he required translation of a sixteen bit 21 logical addres to an eighteen bit physical address is 22 accomplished by a mapping scheme. As part of this 23 mapping scheme, a physical page number is obtained by 24 a look-up operation within a map. ~his physical page 25 number is then combined with the addre~ within a page 26 to form the complete physical memory address~

w 1 Only the page number is transla~ed~ The 2 offset or address within a page is neiJer chanyed in 3 the mapping.

In the present invention there are four 6 map sections. Each map s~ction corre.sponds to one 7 of the four logical addressing areas (user data, 8 system data, user code and system code).

The separation of the logical address into 11 these four separate and distinct areas provides 12 significant bene~its.

14 The separation provides isolation of programs from data so that programs are never modified. The 16 separation also provides isolation of system programs 17 and data from user programs and data, and this pro-18 tects the operating system from user errors.

The four map sections are designated as 21 follows:

23 Map ~--user data map. All addresses to 24 variable user data areas are translated through this user data map.

27 Map l--system data map. The system data 28 map is similar to the user data map and in addition, 29 all memory references by either the I/O channel, the interpxocessor bus handllng microprogram, or the interrupt ` 168 1 handling microprogram specifies this map. The sys~em 2 data map provides channel access to all of physical 3 memory via only a sixteen bit address wor~.

Map 2--user code map. This map defines the 6 active user program. ~11 user instruction~ and constant 7 data are obtained via this user code map.

9 Map 3--system code map. This map defines 10 the operating system program. All operating system 11 instructions and constant data are obtained via ~his 12 system code map.

14 Each map section has sixty-four entries 15 corresponding to the sixty-four pages possible in 16 each logical address area. Each entry contains the 17 following information.

19 (1) The physical page number field ~which 20 can have a value of zero through 255).

22 (2) An odd paxity bit for the map entry.
23 The parity bit is generated by the map logic whenever 24 a map entry is written.

26 ~3) A reference history field. The 27 reference history field compri~es reference bits, 28 and the hi~h order bit of the re~erence bit9 i9 9et 29 to a "one" by any use of the page corre~ponding to 30 that map en~ry.

169 - 5f f''' ~L~l3 1 (4) A dirty bit. The dirty bi~ is set to 2 a "one" when a ~rite access is made to the corr~spond-3 ing memory page.
The reference bits and the dixty bit are 6 used by the memory manager function of the operatiny 7 system to help select a page for overlay. The dirt~
8 bit also provides a way to avoid unnecessary swaps of 9 data pages to secondary memory.
11 (5) An absent bit. The absent bit is 12 initially set to a "one" by the operating system to 13 flag a page as being absent from main m~mory. An 14 access to a page with this bit set to "one" causes 15 an interrupt to the operating system page fault interrupt 16 handler to activate the operating system virtual memory 17 manager function. The absznt bit is also used as a 18 protection mechanism to prevent erroneous access by a `19 program outside its intended logical address area for 20 either code or data.

22 Three instructions are used by the operating _~
23 system in connection with the map. These three ~4 instructions are: SMAP, RMAP, AMAP.
26 The SMAP (set map entry) instruction i9 used 27 by the memory manager ~unction of the operating system 28to insert data inko a map entry. This instruction 2grequires two parame~ers--the map entry address and 30the data to be inserted.

-~.~L3~

1 The ~`~P (read map entxy) instruction is 2 used by the memory manager function of the opcrating 3 system to read a map entry. This instruction requires 4 one parameter, the map en~ry address, and the result returned by the instruction i5 the map entr~ content.

7 Th~ AMAP (ag~ map entry) instruction causes 8 the reference history field of ~ map entry to be shifted 9 one position to the right. This is used by the memory manager function of the operating system to maintain 11 reference history in~ormation as an aid in selecting a 12 page for overlay.

14 A page fault interrupt provided by the absent 15 bit occurs when a re~erence is mad~ to a page that does not 16 currently reside in main memory or which is not part of the 17 logical address space of the program or its data. ~en a 18 page fault is detected, an interrupt through to the operat-19 ing system page fault interrupt handler occurs.

21 The page fault interrupt sequence includes 22 the following events:

24 1~ An address reference is made to a page 25 that is absent ~rom physical memory (absent bit = "one").

27 2. The page ~ault interrupt occurs. The 28 interrupt handler microcode place~ an lnterrupt para-29 meter indicating the map number and the logical page 30 number in a memory location known to the operating ~L3~
~, 1 system. Then the current environment is saved in an 2 interrupt stack mark~r in memory.
4 3. The page fault interrupt handler executés.
5 If the page fault occurred because of a reference out-6 side the logical address space o~ the proqram, then the 7 program is terminated with an error condition. On the 8 other hand, if a page fault occurred because the logical 9 page was absent from physical main memory (but present in 10 secondary memory~, an operating system process ex~cutes 11 to read the absent page from the secondary memory (usually 12 disc) to an available page in primary memory . That 13 physical pa~e information and a zero absent bit are inserted 14 into the map entr~. When this memory management function 15 completes, the environment that caused the page fault is 16 restored.

18 4. The instruction previously causing the page 19 fault is reexecuted. Since the absent bit in the map ~O entry of the logical page has now been set to a "zero", 21 a paye fault will not occur, the page address is trans-22 lated to the physical page just brought in from secondary 23 memory, and the instruction completes.

As ~oted above, the I/O channel has access 26 to the memory through Lt~ own port.

28 Data transers to and rom memory by the I/O
29 channel are via khe system data map. That is, the six-30 teen bit logical addresses provided by the I/O channel 31 are translated to an eighteen bit physical address by 32 means of the system data map.

1 Thus, the mapping scheme allows I/O access 2 to more words of physical memor~ than its a~1dress 3 counter would normally allow.

In one specific embodiment of the present 6 invention 262,144 words of physical memory ~for an 7 eighteen bit address~ can be accessed with only a 8 sixteen bit logical address by going through the map.
g The extra address information (the physical page information) is contained in the map and is supplied 11 by the operating system before each I/O transfer is initiated.

13 As will become more apparent from the 14 detailed description to follow, the present invention is also readily extendible to a twenty bit physical 16 address.

18 Fig. 34 is a bloc'.~ diagram showing details 19 Of the memory 107 of a processor module 33 and showing also connections from the memory 107 to the CPU 105 21 and the I/O channel lOg of that processor module.

23 As illustrated in Fig. 34, the memory system 2~ 107 provides access ports for both the CPU 105 and the I/O channel 109 to the memory 107, and the I/O channel 26 109 therefore is not required to access the memory 27 through the CPU 105.

~g The memory 107 includes map memory control logic 401 which controls initiation and cor.~pletion 31 of access to physical memory modules 403.
` 173 1 The memory 107 also includes a data path 2 section 405 containing registers ~as indicAt~d b'l 3 the legends in Fig. 34 and described in de~ail below) 4 which supply data to be written to memory and which 5 hold data read from memory.

7 The memory 107 also includes a map section 8 407. The map section 407 includes lo~ical address 9 registers from both the CPU and the channel and a map 10 storage 409 from which physical page numbers are obtained.

11 .
12 The map section 407 thus contains a processor 13 memory address (PMA) register 411 and a channel memory 14 address (CMA) register 129.

16 These two registers are connected to an 17 address selector 415.
~ ' .
19 The address selector 415 is connected to the 20 map 409 by a logical page address bus 417, and the address 21 selector 415 is also connected directly to the memory 22 modules by a page offset bus 419.

24 As indicated by the numerals 8 and 10 adjacent 25 to the buses 417 and 419, the logical page address bus 26 417 transmits the eight high order bits to the map 409 27 for translation to a physical page number, and the page 28 offset bus 419 transmits the ten low order bits (of an 29 eighteen page address from the addres~ selector 415) to 30 the memory rnodules 403.
.

~3~

1 An output bus 421 supplies the rJhysic~l paye 2 address to the modules 403. This output bus 421 con~
3 tains the translated eight high order bits f4r the 4 address of the physical page.

6 The data path section 405 contains the follo~7 7 ing re~isters: A processor memory data (PMD) reyister 8 423; a channel memory data (CMD) reyister 425; a next 9 instruction (NI) register 431i a memory data (MD) 10 register 433; and a channel data (CD) register 125.

11 - - , l~ The outputs of the PMD and CMD registers 13 are supplied to a data selector 427. This data 14 selector 427 has an output bllS 429 which supplies data 15 to be written to memory in the modules 403.

17 Data read out from one of the memory modules 18 403 is~reàd into one of the three data registers NI, 19 MD and CD over a bus 437.
21 As illustrated in Fig. 34, the map memory 22 control logic 401 is also connected with each of the 23 memory modules 403 by a bus 439. The bus 439 comprises 24 command lines wh~ch initiate read or write operations, 25 completion signals from the memory modules, and error 26 indicators or flags.

28 With reference now to Fig. 35, the map section 29 407 includes, in addition to the map 409, a map page register 30 441, a map output latch 443, a map memory data (MMD) ~l~6~

1 register 445, a map data selector 447, a map ~arity 2 generator 449, a map parity checker 451, refeLence 3 bit logic 453, and dirty bit logic 455.
The map memory control logic 401 is shown 6 in Fig. 35 as associated with the map section ~07 by 7 control signal lines 457.

9 The map memory control logic 401 controls 1~ the loading of registers and selection of registers by 11 the selectors, controls (in conjunction with map absence 12 and parity error outputs) the initiation of memory 13 modules 403 operations, and provides interrupts to the 14 CPU 105 (as indicated by the page fault and map parity 15 error interrupt signals indicated by the legends in 16 Fig. 35)--all as will be described in more detail below.

18 In a particular embodiment of the invention the l9 memory system shown in Figs. 34 and 35 utilizes a physical 20 page address field of eight bits and a page offset of ten 21 bits which combine to giva a total eighteen bits. As noted 22 a~ove, the numbers 8, 10, 12, 13, 14 and 18 which are not 23 in parenthesis on certain bus lines in Fig. 34 and Fig. 35 24 relate to this specific eighteen bit implemented embodi-25 ment of the present invention. However, the memory system 26 is easily expandable to a twenty bit implemented embodi-- 27 ment (with a physical page address o~ ten bits) and this - 28 is indicated by the numbers (lO), (12), (14), (15), (16) 29 and t20) which are within parenthesis on the same bus 30 lines of Fig. 35.

6~

1 Fig. 36 illustrates the oryanization of 2 logical memory in four separate and distinct logical 3 address areas 459, 461, 463 and 465. These four 4 logical address areas are: user data area 459;
system data area ~61; user code area 463; and system 6 code area 465 8 Fig. 36 also illustrates the four map sections g corresponding to the logical address areas.

11 Thus, the user data map section 467 corres-12 ponds to the logical user data address area 459, the 13 system data map section 469 corresponds to the logical 14 system data address area 461, khe user code map section 471 corresponds to the logical user code address area 16 463 and the system code map sestion 473 corresponds to 17 the logical system code address area 465.

19 As also illustrat~d in Fig~ 3~, each map section has sixty-four logical page entries (page zero 2~ through page sixty-three), and each map entry comprises 22 sixteen bits (as illustrated by the enlarged single 23 map entry in Fig. 36). - ;

As indicated by the legend~ associated with 26 the enlarged map entry shown in Fig. 36, each map 27 entry comprises a ten bit physical page number field, 28 a single parity bit P, a reference history fleld 29 comprising three re~erence bits R, S and T, a single dirty bit D ,and a single absent bit A.

1 The physical page number field pro1~ided by 2 the t~n high order bits provides the physic~l paqe 3 number corresponding to the logical page called for 4 by the program.

S
6 The parity bit P is always gen~rated as odd 7 paxity to provide a data integrity check on the map 8 entry contents.
g The reference history field bits R, S and T
11 are used by the memory manager function of the operating 12 system to maintain reference history information for 13 selecting the least recently used page for overlaying.

The R bit is set ko a one by any read or 16 write operation to that logical page.

18 The S and T bits are storage bits which are 19 manipulated by the ~P (age a map entry) instruction.

21 The dirty bit D is set to a one by a write access 22 to that logical page. The operating system uses the dirty 23 bit to determine whether a data page has been modified 24 since it was last brought in from secondary rnernory.
-26 The absent bit A is set to a one by the operat-27 ing system to flag a logical page which i9 absent from 28 main memory but present in secondary memory or to ~lay 29 a page which is outside the logical address area o~
30 that user.

1~3~
l The two high order bits for the map entry 2 shown in Fi~ . 3 6 are not used in the specific embodiment 3 of ~he invention illustrated in the drawin~s, ~ut these 4 two bits are used when the full twent~ ~it physical S addressing is used.
7 As noted above, three instruc~ions are used 8 by the operating system in connection with the map.
g These three instructions are: SMAP, RMAP and ~AP.

ll The S~P instruction is used by the memory 12 manager function of the operating system to insert data 13 into a map entr~ like that illustrated in Fig. 36.

The SMAP instruction is implemented by the 16 microprogram 115 (Fig. 12) in the CPU lOS. The micro-17 program interacts with the map memory control logic 401 18 ~see Fig. 34), ~irst of all, to select (with the first 19 instruction parameter)a location in the map 409 and then, second, to insert in that location the second 21 instruction parameter--the new map entry data.
22 . ~
2~ In operation, and referring to Fig. 35, in 24 the first step in the sequence the microprogram 115 loads the new map entry data into the processor memory 26 data ~PMD) register 423.
" 27 28 In the next step in the sequence, the map 29 address, including two high vxder blt5 ~or map selection, ;30 are loaded into the processor memory address (PMA) 31 register 411.

: . 17g :~L36~d 1 At this point the two instruction parameters 2 containing the map entry ~ddress and th~ data to be 3 inserted have been lo~ded in their respective regist~rs 4 411 and 423.
6 Mext, the microprogram 115 in the CPU 105 7 initiates a map write operation sequence of the map 8 memory control logic 401. This map write operation 9 sequence is initiated aft~r any previous memor~ operations 10 have been completed.
1~
12 The steps noted above in the operation 13 sequence have all been performed by the microprogram 14 (the firmware).
16 The remaining actions of the SMAP instruction 17 are performed under the control of the map memory control 18 logic. Thus, the remaining actions are all performed 19 automatically by hardware.
~0 ~1 In the map write operation sequence, the map 22 address is transmitted from the P~ register through the 23 address selector 415 over the bus 417 to the map 409.
24 Only the eight high order bits (the map select and map 25 address) are used in this operation.

27 ~he two high order bits speci~y the map 28 Selection--whether user data, system data, user code 29 or system code.

~ 7~

The ten low order bits of the loyical address bus from the address selector (ASEL) ~15 (which bits are the offset within a paye Eor a memory read or write access) are not used in this operation~
As the map is beiny addressed as described above, the new map data is transmitted from the PMD
register 423 through the map data selector 447 to the map parity generator 449 and to the map 409. The map parity generator computes odd parity on the new map data and supplies this parity bi-t to the map.
Now, at this point, the map memory control loyic 401 generates a map write strobe siynal (on one of the lines indicated by 457 in Figure 35) to the map 409 which causes the new data and parity to be written into the selected map section at the specific map entry selected by the loyical page address on the bus 417.
This completes the S~AP instruction se~uence.
At the end of this SMAP instruction the proper map section has been selected, the particular logical page entry on that map section has been selected, the data and computed add parity have been supplied to the map, and the map write strobe has caused that data to be written at the desired map entry.

- 181 ~

1 The SMAP instruction ~SMAP) i5 used by the 2 operating system to initialiY.e each logical page entry 3 in each of the four map SeC~iGnS as required.

One use of the set map instruction is there-5 for to insert a physical page address for a logical 7 page to provide for translation of logical page numbers 8 to physical page numbers after a page has been swapped 9 in from secondary memory.

ll Another use of the set map instruction is to 12 set on an absent bit for a logical page swapped out to 13 secondary memory.

The read map (RMAP) instruction is used by 16 the memory manager function o~ the operating system to 17 examine the content of a map entry.

19 In this RMAP instruction the microprogram 115 20 in the CPU 105 interacts with the map memory control 21 logic 401 to select (with the instruction parameter) a 22location in the map 409 and to return to the register 23stack 112 (see Fig. 12) as a result of the content of 24that map entry.
2~ In the operation of the read map ~RMAP) 27instruction, referring to Fig. 35, the microprogram 28115 loads the map address, including the two high order , 1~3~Z~

1 bits for the map selectionl into the P~A register 411.
2 The microprogram 115 then initiates a map read operation 3 se~uence of the map memory control lcgic 401 This sequence is th~n carried out b~ the hard-6 ware, and in this sequence the ma~ address is transmitted 7 from the PMA register 411 through the address selector 8 415 to the map 409. Again, only the map select and page g address bits are used in this operation.
10 .
11 The content of the selected map entry is 1~ transmitted from the map 409 to the map parity checker 13 451 (see Fig. 35) and to the map output latch 443. The 14 map parity checker 451 compares the parity bit ~rom the map entry with the odd parity computed on the data.

17 If the parity is incorrect, the map address 1~ is loaded into the map page register 441; and the map 19 parity error signal sets an error ~lag which causes a 20 map parity error interrupt to the CPU 105.

22 Otherwise, in the case of correct parity, 23 the map entry data is loaded from the map output latch 24 443 into the map memory daka registex (MMD) 445, 26 Finally, the,RMAP instruction micrGprogram 27 returns ~he data in the map memory data (MMD) regi~ter 28 445 to the register stack 112 (see Fig. 12) a~ the result 29 Of the instruction.

!

183 , I

. ~

1 At the end of the read map (RM~sP) ins~ruction 2 the proper map section has b~en selected, the par~icular 3 logical page entry on that map section has been selected, 4 and the content o~ that map entry has been read out from the map and returned as an instruction result to the CPU's 6 register stack.
8 The uses of the RMAP instruction include the 9 following.

11 Th~ main function of this read map ~RMAP) 12 instruction is to allow the operating system to examine 13 the re~erence history field and dirty bit of a map 14 entry tsee the map entry format shown in Fig. 36) to 15 determine a.page for overlaying (as will become more 16 apparent from the description of the operation to follow).

18 The read map ~RMAP) instruction is also used 19 in diagnostics to determine whether the map storage is 20 functioning properly.

22 The age map (AMAP) instruction is used by the 23 memory manager Xunction of the operating system to maintain ~4 useful reference history in~ormation in the map. This 25 reference history information is maintained in the map 26 by map entrieg tthe R,Sand T bits of the map entry 27 format shown in Fig. 36) within a map section which are 28 typically "aged" after each page fault interrupt occurrence 29 in that map section.

1 This AMAP instruction has ju~t a single para-2 meter which is the map address specifying the map 3 location to be aged~

In the operation of the age map (I~P) 6 instruction, the microprogram 115 in the CPu 105 selects 7 a map location with the instruction map address parameter.
8 The microprogram 115 loads the map address parameter into 9 the PM~ register just as in the RMAP instruction.

11 At this point a map--read operation sequence 12 Of the map m~mory control logic 401 is initiated, and 13 this sequence proceeds identically as in the ~ ~P
14 instruction described above.
1~ .
16 The microprogram 115 (Fig. 12) reads the 17 content of the map entry from the MMD register 445 (Fig.
18 35) extracts the reference history field (the R, S and T
19 bits lO, ll and 12 5hown in Fig. 36), shifts the field 20 right one position, and reinserts the field to form the 21 new ma~ entry data. Thusr a zero has been entered in 22 the R bit, the R bit has been shifted into the S bit, 23 the S bit has been shifted into ~he T ~it, and the old 24 T bit is lost.
26 Now the microprogram 115 takes the modi~ied 27 map entry and loads this new data into the PMD register ~8 ~23 (Fig. 34) and writes the new map entry data back into 29 the selected map entry (similar to the S~AP sequence).

~L~L3~7~

1 This completes the age map (A~IAP) ~ instruction.

4 As a result of the age map (AMAP) instruction, a map entry has been read from the map, itC reference 6 history field has been shifted, and this modified entr~
7 has been reinserted into the selected map location.

9 As previously noted, the R bit is set to one 1~ by any memory reference to the corresponding logical page, 11 so that when this bit is a one, it is an indication 12 that this page has been used since the last set map 13 (SMAP) or age map (AMAP) operation instruction.

15 This setting of the R bit in conjunction with 16 the age map (~MAP) instruction provides a means for 17 maintaining frequency of use information in the reference 18 history field of the map.

20 The reference history ~ield of all of the map ?
21 entries in a given map are typically aged after a page -~2 fault interrupt. Thus, the value of the three bit 23 reference field in a map entry is an indication of the ~4 frequency of access since the previous three page fault t 25 interrupts.

27 For example, a binary value of seven (all 28 three reference bits set at on~), indlcates accesses in 29 each of the intervals between the proceeding page fault 30 int~rrupts.

f ~36~

1 A binary value of four in the reference history 2 field (the R bit set at one and the S and T bits set 3 at zero) indicates an access in the interval since the 4 last page fault interrupt and indicates that there were S no accesses in the intervals previous to the most recent 6 page fault interxupt.
8 As a final example, a binary value of zero for 9 the three bit reference field indicates that that logical 10 page has not been accessed in any of the three intervals 11 since the last three page fault interrupts.
~2 13 ~hus, the higher the binar~ number represented 14 by the three bit reference history field, the higher the 15 frequency of recent accesses to that logical page.

17 . ThiS reference history information is main-18 tained so that whPn it is necessary to select a page for 19 overlay, a page which has been infrequently used in the 20 recent past can be identified. A page infrequ~ntly 21 accessed in the recent past is likely to continue that 22 ~ehavior, and that page will therefore probably not have 23 to be swapped back into memory after being overlayed.

This frequency of u~e history is uqed by the 26 memory manager function of the operating 5ystem to select 27 infrequently used pages for overlay so as to minimize 28 swapping from secondary memory to implement an efficient 29 virtual memory system.

3L~3~

1 As noted above, memory may b~ accessea by 2 the CPU or by the I/O system.

4 The action of the memory system and map during S a CPU memory access sequence will now be described.
6 The access sequence is similar for the various CPU
7 memory accesses such as writing data, reading data, or 8 reading instructions from memory.

1~ The CPU memory access sequence is started 11 either by the CPU microprogram 115 or by the CPU
12 instruction-fetch logic. In either event, the CPU 105 13 loads an eighteen bit logical address into the PMA
14 register 411 and initiates a data read, data write, or ~5 instruction r~ad operation sequence of the map memory 16 control logic 401.

18 The eighteen bit logical address is composed 19 of two high order logical address space select bits and 20 sixteen low order bits specifying a location within that 21 logical address space. The two select bits may be 22 specified by the CPU microprogram 115 or may be auto-23 matically generated in the CPU, based on the contents of 24 the instruction (I) and environment (E) registers.

26 The eighteen bit logical address also includes, 27 in addition to the two high order logical address select 28 bits, six bits which specify the logical page within the 29 selected map and ten low order bits which speci~y the 30 offset wi~hin the page in the selected map.

~r ~13~

1 In the data read, data ~/rite, or instructi~n 2 read operation sequence of the map memor~ corltrol logic 3 401~ after any previous map or memory operations have 4 completed, the eighte~n bit address in the ~A register 411 (Fig. 35) is transmitt~d through the addr~ss selector 6 415 to the bus~s 417 and 419 (see Figs. 34 and 35~.

8 The bus 419 transmits the page orfset portion g of the address. This page offset portion of the address is transmitted dir~ctly to the physical memory modules 11 403 (Fig. 403) by the bus 419.
~2 The bus 417 transmits the logical page address 14 portion (which must be translated to a physical page address) to the map 409.

17 The map entry selPcted by the logical page 18 address is read out from the map 409 to the map memory 19 control logic 401 (Fig. 34~, the map parity checker 451 (~ig. 35), and the map output latch 443.

22 . I~ the absent bit is a one, the logical page 23 address is loaded into the map page register 441, a 24 page fault interrupt signal is transmitted tv the CPU
25 105, and the map memory control logic 401 terminates 26 the memory access sequence.

28 Similarly, if the parity checker 451 detects 29 incorrect parity in the map entry, the logical page 30 address is loaded into the map page register 441, a map ~1~6~ 8 1 parity error signal is transmitted to the CPU, and the 2 memory access sequence is ~erminated.

4 Otherwise, if there is no error, the physical 5 page address is transmitted from the map output latch 6 443 over the bus 421 to the physical memory modules 403;
7 and the map memory control logic 401 issues a command 8 over the bus 439 to cause the selected memory module 403 9 to perform a read or write operation.

11 In a CPU write operation the data to be written 12 is transmitted from the PMD register 423 through the 13 data selector 427 to the memory module over the bus 429.

While the memory module is performing a read 16 or write operation, the map memory control logic 401 17 causes,the map entry data to be modified and rewritten.

19 The map entry data, without the parity bit 20 P.or the reference bit R, is transmitted from the map 21 output latch 443 to the dirty bit logic 455 ~see Fig.
22 35) and to the map data selector 447.

24 In this operation the physical page field 25 of a map entry (shown in enlarged detail in the lower 26 righthand part o Fig. 36) and the S and T bits o the 27 reference ~ield and the ab~ent bit are alway5 rewritten 28 without modification.

' 190 ~l13S~
1 If a CPU data write oper~tion is being 2 perform~d, the dirt~ bi~ D su~pli~d to th~ map data 3 selector is set to a one by the dirty bit logic 455.
4 Otherwise, the dirty bit is not modified.

6 The rPference bit ~ supplied to the map data 7 selector by the reference bit logic 453 is set to a one 8 in either a read or a write operation.

The phy~ical page field and the S, T and A
ll bits are not modified, as noted above.

13 The map data selector 447 supplies this new 14 map data to the parity generator 449 and to the map 409.
lS
16 An odd parity bit P is generated ~rom the 17 new data by the parity generator 449 (see Fig. 35).

19 A map write 3~robe from the map memory control 20 logic 401 then causes the new data and parity to be 21 written intv the map entry selected by the logical page 22 address bus 417.

24 ~hus, the logical page has been translated 25 through the map entry, and the map entry has been 26 rewritten with updated parity, reference,and dirty bits.

28 When the physical memory module 403 completes 29 its read or write operation, it sends a completion signal 30 to the map memory co~trol logic 401 over the bus 439 31 (see Fig. 34).

.... . ..

3L~3~

1 In a read operation the mcmory module 403 2 gates the memory data to the bus 437 (Fig. 34~.

4 In a data read operation sequence the data 5 is loaded into the MD register 433 (Fig. 34) for use 6 by the CPU lOS.

8 Xn an instruction read operation sequence g the data is loaded into the NI register 431 (Fig. 34) for subsequent execution by the CPU 105.

1~ ., .
12 The C~U memory accesses of data read, data 13 write and instruction read are thus completed as 14 described above.
16 An I/O channel access to read or to write 17 data to memory proceeds similar to a CPU memory access 18 as des~ribed above except for the following.

The channel memory address (CMA) register 21 129 (Fig. 34) is used to provide the logical address, 22 and this register always specifies the system data map 23 469 ~see Fig. 35).

The channel memory data (CMD) register 425 26 (Fig. 34) is used to supply data to memory in a write 27 Operation.

29 The channel data (CD) register 125 (Fig. 34) is used to recelve data from memory in a read operation.

1~2 1 In an I/O channeL lO9 merrory acc~ss, the 2 access is always a read or write data to memory access, 3 and there is no instruction read access as in the case 4 of a CPU access.
6 In addition, map parity and absent conditions 7 are transmitted to the I/O channel lO9 i~ they occur in 8 an I/O channel access to memory.

As noted at several points above, either ll semiconductor memory core memory is used for the memory 12 modules 403.

14 When the memory is core memory, errors are 15 detecteZ by a parity error detection system. The parity 16 error detection system for core memory modules is 17 effective to detect all single bit errors. Conventional 18 parity error generation and checking techniques are used, 19 and details of the core memory will therefore not be 20 illustrated.

22 The probability of failures in semiconductor 23 memory is great enough to justify an error detection 24 and correction system, and the present invention provides 25 a detection and coxxection system which incorporates a 26 six ~it check field for each sixteen bit data word. Figs.
27 37-41 and ~elated Table 1 (set out below) illustrate 28 details of an error detection and correction system used 29 when the memory modules 403 are constructed with semi-30 cond~ctor memory.

Lg3 ~-~ . .... .

The six bit check field error detection and correction system of the present invention is, as will be described in detail below, capable of detectiny and correcting all single bit errors and is also capable of detecting all double bit errors. In addition, most errors of three of more bits are detected.
While the error detection and correction system will be described with re~erence to a semi-conductor memory, it should be noted that the system is not limited or restricted to semiconductor memory but is instead useful for any data storage or trans-mission application.
An important benefit of the error detection and correction system o~ the present invention results from the fact that not only are sinyle bit errors corrected but also that any subsequent double bit errors are reliably detected after a single bit has failed.
The multiprocessor system incorporating the error detection and correction system of the present invention is therefore tolerant of single failures and can be operated with single bit failures in semiconduc-tor memory until such time as it is convenient to repair the memory.
The error detection and correction system utilizes a systematic linear binar~ code of Hamming dis-tance four. In this code each check bit i8 a linear combination of eight data bits (as shown in Figure 38).

- 19~ -r ~3~7~

1 Also, each data ~it is a component of exactly three 2 check bits (as also shown in Fiy. 38). ~n advan~age of 3 this code is that uniform coverage of the data bits 4 by the check bits i5 obtained.
6 The error correction and detection system 7 embodies a syndrome decoder which provides the 8 combination of fast logic speed and low parts count.

In initial summary, the error detection and 11 correction system of the present invention operates to 12 add six check bits to each data word written into stor-13 age. When a data word is subsequently read out o~ memory, 14 the check field portion of the storage word is used to identify or to detect the loss o~ information in that 16 word since the tLme it was stored.

18 ~ In semiconductor memory there are two possible 19 mechanisms for loss of information (error). One is hard 20 failure of a memory device which makes that device 21 permanently unable to retain information, and the other 22 is soft failure in which electrical noise can cause a 23 transient loss of information.

The detection o~ error.s is accompllshed by a 26 check bit compar~tor which produces a ~ix bit s~ndrome.
27 The syndrome is the difference between the check field 28 obtained from the stored word and the check field which 29 would normally correspond to the data ~ield obtained from 30 the stored word.

~3~7~

1 This syndrome is then anal~/zed (decoded) to 2 determine whether an error has occurred and, if an error 3 has occurred, to determine wha~ t~pe of correction is re~uired.

S In the case of single data bit errors, the 6 syndrome decoder output causes a data ~it complementer 7 to invert the bit that was in error; and this corrected 8 data is supplied as the output of that memory rnodule.

If the syndrome decoder indicates a multiple 11 error, then the fact of the multiple error is communicated 12 to the map memory control section by means of one of 13 the control and error lines to cause an interrupt to the CPU.

lS With reference now to Fig. 37, the memory 16 module 403 includes a timing and control logic section 17 475 and a semiconductor storage array 477. The storage 18 array 477 provides storage for 3Z,768 words of twenty-19 two bits each. Each word has (as illustrated in Fig. 37) 20 a sixteen bit data field and a six bit check field.

~2 Each semiconductor memory module 403 also 23 has, as illustrated in Fig. 37, an output latch 479, 24 a check bit generator 48i, a check bit comparator 483, 25 a s~ndrome decoder 485 and a data bit complemen~er 487.

27 The memory module 403 interfaces to the rest 28 of the system through the signal and data paths illustrated 29 in Fig. 37. These paths include: 429 (data to memory bus), 30 439 tcontrol and error lines to the map memory control 1 section 401), 419 and 421 (physical addr~ss bus), ~nd 2 437 (data from memory bus). These signal and data pa~hs 3 are also shown in Fig. 34 With continu~d reference ~o F~ig. 37, the 6 content of the output latch 479 is transmitted on a 7 bus 489 to both the check bit comparator 483 and the 8 data bit comparator 487.

The output of the check bit comparator 483 is 11 transmitted on a syndrome bus 491 to both the s~ndrome 12 decoder 485 and the timing and control logic section 475.

14 The output of the syndrome decoder 485 is trans-15 mitted on a bus 493 to the data bit complementer 487.

17 . Other outputs o the syndrome decoder 485 18 are transmitted on lines 495 and 497 to the timing and 19 con~rol logic section 475. The line ag5 transmits a 20 SINGLE ERROR (correctable error) signal, and the line 21 497 transmits a MULTIPLE ERROR (uncorrectable error) signal.

23 The timing and control logic 475 provides 24 control signals on a control bus 499 to the semi-25 conductor storage array 477 and also to the output latch 26 479.

28 The output o~ the check bit generator 481 is 29 transmitted to the storage arra~ 477 by a bus 501.

lg7 ~3~7~3 With reference to Fiyure 38, the check bit generator 481 includes six separate eiyht-bit parity trees 503.
As shown in Figure 39, the check bi-t compara-tor 483 includes six separate nine-bit parity trees 505.
As shown in Figure 40, the syndrome decoder 485 includes a decoder section 507 and a six-bit parity tree 509~
With continued reference to Figure 40, the outputs of the decoder section 507 and six-bit parity tree 509 are combined in error identification logic indicated generally by the reference numeral 511.
As illustrated in Figure 41, the bit com-plementer 437 comprises sixteen exclusive-or gates 513.
In operation the sixteen bit data word is sup-plied by the bus 429 to the storage array 477 and also to the check bit generator 481 (see Figure 37).
The check bit generator 481, as best illustrated in Figure 38, generates six check bits C0 through C5 by means of the six eight-bit parity trees 503.
As also illustrated in Figure 38, the eight-bit parity tree 503 farthest to ~he le~t generates ,~
' ~ ' ~.~36~

l check bit zero (C0) as sp~cifi~d by th~ lo~Jic ecluation 2 for C0 as set out at the lower part of Yig. 38. Check 3 bit zero (C0) is therefore the complement of the modulo-4 two sum of data bits 8 through 15.
6 By way of further example, the chec~. kit C3 7 is generated by an eight bit parity tree 503 as specified by the logic equation for C3 set out at the lower part 9 of Fig. 38. Check bit three (C3) is the modulo-two 10 sum of data bits 0, l, 2, 4, 7, 9, 10 and 12 as shown 11 by the logic equation and as also illustrated by the 12 connections between the eight bit parity tree and the 13 corresponding data bit lines in the logic diagram in the 14 upper part of Fig. 38.

16 Similarly, each of the other check bits is 17 generated ~y a modulo-two addition of eight data bits 18 as illustrated in the logic diagram in the ~op part of 19 Fig. 38.
21 To accomplish a memory write operation, these 22 six check bits, as thus generated by the check bit 23 generator 481, and the sixteen data bits, as transmitted 24 on the data bus 429, are entered in a particular location 25 in the storage array 477. As illustrated in Fig. 37, the 26 six check bits and the sixteen data bits are entered 27 in the storage axray 477 under the control of the timing 28 and control logic 475 and the physical address i~formation on 29 the physical address bus 419, 421.

.
lg9 . . -- .. . . . .. ...

1 Every word stored in the storage array 477 2 has a six bit check field genera~ed for that ~lord in 3 a similar manner. This check field is retained with 4 the stored word in the storage array ~77 until ~he time 5 when that location in the storage array is subsequently 6 accessed for a read operation.

8 When a particular word is to be read out o~
9 the storage array 477, the timing and control logic 475 10 and the address on the physical address bus 419, 421 11 causes the content of the selected s~orage location to 12 be loaded into the output latch 479. The output latch 13 is twenty-two bits wide to accommodate the sixteen data 14 bits and the six bit check field.

16 From the output latch 479 the sixteen data 17 bits and the six bit check field are transmitted by a 18 bus 489 to the check bit comparator 483.

As illustrated in Fig. 39, the check bit 21 comparator 483 forms six syndrome bits S0 through S5.

23 Each syndrome bit is the output of a nine-bit 24 parity tree 505 whose inputs are eight data bits and one 25 check bit. Each syndrome bit is related to a cvrrespond-26 ingly numbered check bit. Thus, check bit zero is used 27 only for computing syndrome bit zero, check bit one is 28 used only for computing syndrome bit one, and so ~orth.

. .

. . .

~L3G~

As an example, syndroTne bit zero (S0) is the complement of the modulo-two sum of check bit zero and data bits 8 through ].5 (as shown in the logic equation at the bottom of Figure 39)~
Similarly, each of syndrome bits S 1 throuyh S 5 is generated from the modulo two sum of a corres-ponding check bit and eight of the data bits, as shown by the connections to the particular data bit llnes for each syndrome bit in the logic diagram part of Figure 39.
The presence or absence of errors and the types of errors, if any, are identified by interpreting the value of the six syndrome bits on the bus 491.
Table l enumerates the sixty-four possible values of the six bit syndrome code and gives the inter- .
pretation for each possible value. ~.

-~ 3~72&~

TABLE I
SYNDROME CODES
S0 Sl S2 S3 S4 S5 ERROR INS0 Sl S2 S3 S4 S5 ERROR IN
O O O O O O (No Error)1 0 0 0 C0 O O O O O 1 C5 0 0 0 1 ~Double) O O 1 0 C4 0 0 1 0 ~Double) O ~ Double) O O 1 1 D8 O 1 0 0 C3 0 1 0 0 ~Double) 0 1 0 1( Doub 1 e) O 1 0 1 D9 0 1 1 0 (Double) O 1 1 0 D10 O O O 1 1 1 D0 1 1 1 (Double) O O 1 0 .0 0 C2 1 0 1 0 0 0 (Double) O O O 1 (Double~ O O O 1 Dll O O 1 0 (Double) O O 1 0~Multi-All O's) O O 1 1 ~Multi) O O 1 1 (Double) O 1 0 0 ~Double) O 1 0 0 DlZ
O 1 0 1 D]. O 1 0 1 tDouble) O 1 1 0 D2 0 1 1 0 ~Double) : O 1 1 1 ~Double) O 1 1 1 tmulti) . 20 0 1 0 0 0 0 Cl 1 1 0 0 0 0 ~Double) O O O 1 (Double) O O O 1 D13 ~; O O 1 0 (Double) O O 1 0 D14 O O 1 1 D3 0 0 1 1 (Double) :- O 1 0 0 (Double) O 1 0 0 (Multi) O 1 0 1(~lulti-All l's) O 1 0 1 (Double) O 1 1 0 D4 0 1 1 0 (Double) O 1 1 1 (Double) O 1 1 1 (Multi) O 1 1 0 0 0 (Double) 1 1 1 0 0 0 D15 O O O 1 D5 0 0 0 1 (Double) - 30 0 0 1 0 D6 0 0 1 0 (Double) " O O 1 1 (Double~ O O 1 1 (Multi) ` O 1 0 0 D7 0 1 0 0 (Double) O 1 0 1 (Double) O 1 0 1 (Multi) O 1 1 0 (Doublo) O 1 1 0 (Multi) O 1 1 1 (Multi) O 1 1 1 ~Double) THUS (NUhlBER OP 1 ' s IN ~SYNDRO~lE~
~ 0 BITS - NO ERROR 3 BI'l'S - DATA BIT OR MULTI
; 1 BIT - CHECK BIT ERROR 4 BITS - DOUBLE
2 BITS - DOUBLE 5 BTTS - ~IIJLTI

~ ~ 3 6~ Z~

1 For example, i~ all of the s~ndrome bits S 0 2 throuqh S , are zero, there i5 no error in either th~
3 data field or the check field. This is the condition 4 illustrated at the upper left of Table 1.
s 6 The presence or absence of errors and ~he 7 type of error is su~arized at the botto~ of Table 1.

9 In this summarization, ~Ihen all six syndrome 0 bits are 2ero, there is no error, as noted above.

~3~j7;2~
1 If only one of the six syndrome bits is on, 2 this indicates an error in the corresponding check bit.
3 It should be noted at this point that check bit errors 4 are single bit errors which do not require correction 5 of the data word.

7 As also illustrated in the summary at the 8 bottom of Table l, when two bits are on there is a ~ double bit error; and the two errors could be ~a) one 10 error in a data bit and one error in a check bit or 11 (b) two errors in the data bits or ~c) two errors in 1~ the check bits.

14 When three bits are on in the six bit syndrome 15 code, that condition can correspond to either a single 16 data bit error or a ~,ultiple error.

18 As an example of a single bit error in a data 19 bit, see the syndrome code 111,000 indicating a single 20 bit error in data bit D-15 in the lower right hand part 2l of Table 1. As will be described in more detail below, 22 the syndrome decoder 485 (Fig. 37 and Fig. 40~ will cause 23 the incorrect value of data bit 15 to be inverted 24 (corrected).

26 The ~yndrome decoder 485 provides two ~unctions.
27 First, the syndrome decoder 485 provides an 28 input to the data bit complementer 487 (see Fig. 37) by 29 way of the bus 493 in the case of single data bit errors, 3~ which input causes the erroneous bit to be inverted with-31 in the data bit compl~menter 487.
204 _;

~L3~

1 Secondly, the syndrome decoder 485 provides 2 one of two error signals in the event o~ an error.

4 A single data or check bit error is transrnitt~d 5 on the SINGLE ERR~R line 495 ~o the timiny and control 6 logic 475.

8 A multiple error indication is transmitted on g the MULTIPLE ERROR line 497 to the timing and control 10 logic 4js, 12 A MUL~IPI~E ERROR signal is generated in the 13 case of all double bit errors and most thr~e or more 14 bit errors. This MULTIPLE ERROR signal, as noted above, 15 causes an interrupt to the CPU 105 (see Fig. 34).
17 The construction of the syndrome decoder 485 18 is shown in detail in Fiy. 40. The syndrome decoder 485 19 comprises a decoder 507, a six bit parity tree 509 and 20 err~r identification logic 511.

22` The decoder 507 decodes five of the six syndrcme 23 bits (bits Sl through S5) to provide sufficient information 24 (thirty-two outputs) to generate both the erxor types 25 (whether singla errors or double or multiple errors) and 26 the sixteen output lines required or inver~ion o~ data 27 bit errors in the sixteen data bits. These ~ixteen output 28 lines required for inversion of data bit errors are 29 indicated generally by the bus 493 and are identified 30 individually by T0 through T15 in Fig. 40.
.

1 The decoder 507 output5 which are not connected 2 to the OR gate 512 correspond to errors in the six check 3 bits. Errors in the six check bits do not need to be 4 corrected (since the errors are not data bit errors), 5 and these outputs of the decoder ar~ therefore not us~d.

7 The remaining outputs (the outputs connected 8 to the OR gate 512~ represent double or mul~iple errors g and are so indicated by the legends in Fig. 40. All of 10 these cases are collected by the O~ gate 512 and are one 11 component of the multiple error signal on the line 497 12 at the output of the error identification logic 511.

14 As also il.lustrated in Fig. 40, the syndrome decoder ~5 includes a parity tree 509 which forms the 1~ modulo-two sum of syndrome bits S0 throush 55.

18 ~he resulting even or odd output of the parity 19 tree 509 corresponds to the error classes shown at the 20 bottom of Table 1.

22 Thus, the EVEN output 514 corresponds to syndromes 23 containing no bits on, two bits on, four bits on, or six 24 bits on.

26 ~he EVEN syndrome corresponding to no bits on 27 ~no error) is excluded from the MULTIPLE ERROR output signal 28 497 by an AND gate 515 which excludes the æero syndrome 29 case (the other input from decoder 507 to the gate 515).

~L~L3~

1 Sy~dromes containing two bit~ on, four ~its on 2 or six bits on are thus the o~ly remaininc3 ~ N s~ndromes 3 which in combination with the MULTIPL~ signal constitute 4 multiple errors as transmitted on the output line MULTIPLE
5 ERROR (497).

.7 An output is desired on the SINGLE ERROR
8 indicator line 495 only for single bit errors. Since 9 the odd output on the line 510 of the parity tree 509 1~ corresponds to one bit on (check bit error), three bits 11 on (data bit error or multibit errors), or five bits on 12 (multibit-errors) in the six~bit syndrome (as indicated in 13 the summary at the bottom of Table 1), the odd output on 14 line 510 must be quali~ied so that only single bit errors 15 are transmitted through the logic 511 to the line 495.
16 Those three-bit syndrome codes corresponding to multibit 17 errors and all of the five-bit syndrome codes must there-18 fore be excluded so that onl~ the single bit errors are 19 transmitted on the line 495. This is accomplished by an ?0 inverter 517 and an AND gate 519.

~2 A SINGLE ERROR output is generated on the line ~3 495 for syndrome codes containing ~ single one bit (check 2~ bit errors) and also or those syndrome codes containing 25 three one bits corresponding to data bit errors. As noted 26 above, the odd output of the parit~f tree 509 indicates 27 syndromes containing one, three or ~ive bits on. The 28 inverter 517 and the AND gate 519 exclude multiple error 29 three bit syndromes and all ~ive bit syndromes. Thus, 30 the SINGLE ERROR output 495 includes only single check 2~07 bit errors and sinyle data bit errors. Sinyle check bit errors do not need to be corrected, and single data bit errors are corrected by the bit complementer 487.
The loyic equations for M~LTIPLE ERROR and for SINGLE ERROR listed on the bottom of Fiyure 40 represent the operation described above.
There are some errors of three or more bits which are not identified as multiple errors and in fact can be incorrectly identified as no errors or as single bit errors (correctable errors). However, the normal pattern of error generation is such that the deteriora-tion of storage is normally detected before three bit errors occur. For example, the normal pattern of deterioration of memory storage would first involve a single bit exror from noise or component failure, then would later involve a double bit error from additional failure, etc.; and the double bit errors would be detect-ed before the three or more bit errors could develop.
The function of the data bit complementer 487 (see Figure 37) is to invert data bit errors as detected by the syndrome decoder 485.
Figure 41 shows details of the construction of the bit complementer 487. As illustrated in Figure 41, the bit complementer 487 is implemented by exclusive-or gates 513. Each of these yates 513 inverts a given data bit on a line 489 when a corresponding decoder output on a line 493 is asserted.

3L~L3~7Xi~

1 The corrected output is then ~ransmitted on an 2 output line 437 of the bit c~mplzmenter 4~7 as the out-3 put of that physical ~emory module.
This completes the descri~tion of the error 6 detection and correction system.
.

8 The memory system of the present invention g provides a number of significant features.
11 First of all, the memory map provides four 12 separate and distinct logical addrPss spaces--system 13 code, s~stem data, user code and user data--and provides 14 for a translation of logical addresses within these 15 address spaces to physical addresses.

17 The division of logical memory into four 18 address spaces isolates the system programs from the 1~ actions of the user programs and protects the system 20 programs from any user errors. The division into four 21 logical address areas also provides for a separation of 22 code and data for both user code and data and system 23 code and data. This provides the benefits of non-24 modifiable programs.
26 There are specific ~ields within each map 27 entry for this page address translation and for other 28 specific conditions.
~9 One field permits translation of logical page 31 addresses to physical page addresses.
~09 ~3~

1 Another field provides an absencc indic~tion.
2 This field is an absence bit which allows implementation 3 of a virtual memory scheme where logical payes may reside 4 in a secondary memory.

6 Another field is a reference history field.
7 This reference history field allows frequency of use 8 information to be maintained for use by the memory manager 9 function of the operating system to make the virtual memory scheme an efficient scheme. Frequently accessed 11 pages are retained in primary memory, and infrequently 12 used pages are selected for necessary overlaying.

14 A dirty bit field is maintained in each entry 15 of the system data map and the user data map so that 16 unmodified data pages can be identified. The unmodified 17 data pages so identified are not swapped out to secondary 18memory because a valid copy of that data page is already ~9present in secondary memory.

21 The memory system includes map memory control 221Ogic which automatically maintains the reference and ~3dirty bit information as CPU and I/O channel accesses 24are made to memory.

26 The memory system of the present invention pro-27vides for three CPU instructions--SMAP, RMAP and AMAP~-2gwhich are used by the operating system's memory manager 2gfunction to maintain and to utllize information in the 3~ap-~ ,, ~L3~7~3 1 The memory system of the present invention2 includes a dual port access to the memory. The m~mory 3 can be accessed separately by the CPU and by the I/O
4 channel. Accesses to memory by the I/O channel do 5 not need to involve the CPU, and the CPU can be per-6 forming other functions during the time that an I/O
7 data transfer is being made into or ou~ of memory.

9 The operation of the dual port access to 10 the memory also involves arbitration by the map memory 11 control logic in the event that the CPU and the I/O
12 channel attempt a simultaneous access to the memory.
13 In the case o~ simultaneous access, the I/O channel 14 is given priority and the CPU waits until that particular 15 I/O channel access has completed.

l? Physical memory is expandible by the modular 18 addition of physical memory modules.

The physical memory modules incorporate, in 21 the case of semiconductor memory, error detection and 22 correction under certain conditions. Single errors 23 are detected and corrected so that operation of the 24 CPU and I/O channel can be continued even in the event 25 Of a transient or permanent failure within the phy~ical 2~ memory module. The error detection and correction 27 system comprises a twenty-two blt word within the storage 28 medium. Sixteen bits represent the data and six bits 29 provide an error detection and correction check ~içld.
30 The six bit check fiéld allows the detection and f 1 correction of all single errors and the detection of 2 all dou~le errors.

4 The core memory includes parity for the 5 detection of single errors.

7 - In the overall multiprocessor system of th~
8 present invention each processor module incorporates its g own primary memory system.

11 Since each processor module has its own memory 12 system, problems o~ shared memory in a multiprocessing 13 SyStem are avoided.

The problems of shared memory in a multiprocessing 16 system include reduced memory bandwidth available to a 17 particular processor because of contention, and this 18 reduction o~ available memory bandwidth becomes more 19 severe as additional CPU's are combined with a single 20 shared memory.

22 The probl~ms of interlocks relating to the 23 communication between CPU's by means of areas within a 24 shared memory are avoided by the present invention which 25 does not include shared memory and whlch does, instead, 26 provide for communication between procesqor modules by 27 an interprocessor bus communication system.

2g An additional problem of shared memory is 30 that a failure in the shared memory can result in 3~

1 simultaneous failure of some or ~11 of th~ C~U's in 2 the system. That is, in a shared m~mor~ s~stcm, ~
3 single memory failure can stop all or part of the sys~em;
4 but a memory failure will not stop the multiprocessor 5 system of the present invention, 7 The dual port access by the CPU and the I/O
8 channel to the memory utilizes and is permitted because of separate address registers and separate data registers 0 to and from memory.

12 The CPU has a specific register (the NI register) 13 specifically for receiving instructions frGm memory. This 14 separate and specific register allows overlapped fetch-15 ing of the next instruction duriny execution of the 16 currer.t instruction (which may involve the reading of 17 data from memory). As a result, at the end of a current 18 instruction, the next instruction can be initiated immedi-l9 atelv without waiting for an instruction re~ch.
~0 2~ The map is constructed to provide significantly 22 ~aster access than the access to physical main memory, 23 This provides a number of benefits in the translation of 24 addresses through the map, 26 As one result, in the memory ~ystem o~ the 27 present inVention, the map can be xewritten ~n the time 28 that the physical memoxy acces6 is being accomplished.

3~

~i3~

Because the rewriting is so fast, the rewritin~
of the map does not increase memory cycle time.
Also, the high speed at which the map can be accessed reduces the overall time including page trans-lation required for a memory access.
Parity is maintained and checked in the actual map storage iself. This provides immediate indication of any failure in the map storage before resulting in-correct operation in the processor module can occur.

Claims (5)

THE EMBODIMENTS OF THE INVENTION IN WHICH AN EXCLUSIVE
PROPERTY OR PRIVILEGE IS CLAIMED ARE DEFINED AS FOLLOWS:

1. A multiprocessor system of the kind in which separate processor modules operate concurrently and cooperatively for system control and application processing, said multiprocessor system comprising, a plurality of separate processor modules, interprocessor bus means including an interprocessor bus connecting each processor module for direct processor module to processor module signaling and data transfer and wherein the interprocessor bus means are separate and distinct from an input/output system and from a memory system bus, each processor module having a central processing unit, a memory for that module, and interprocessor control means for connecting the processor module to the interprocessor bus for signaling and data transfer with another processor module, bus controller means operatively associated with the interprocessor bus and each interprocessor control means for determining the priority of data transfers between any two processor modules over the interprocessor bus and for controlling transmissions over the interprocessor bus, each processor module including input/output channel means for transferring data between the processor module and one or more device controllers, a plurality of peripheral devices, a plurality of device controllers for controlling the transfer of data between the processor modules and the peripher-al devices, a plurality of ports in each device controller, a plurality of input/output buses connecting each device controller for access by a plurality of different pro-cessor modules and wherein the input/output channel means, input/output buses and device controllers are operatively connected to pro-vide, directly or in combination with the interprocessor bus means and a processor module, a path for data transfer between each processor module and any device controller.

2. The invention defined in claim 1 including a second interprocessor bus interconnecting each processor module and providing an alternate path for direct processor module to pro-cessor module signaling and data transfer.

3. The invention defined in claim 2 including distributed power supply means for insuring continuous operation of the remainder of the multiprocessor system in the event of a failure of a single power supply to a part of the multiprocessor system, said distributed power supply means including a separate power supply for each processor module and connecting means for con-necting a plurality of the separate power supplies to each device controller and effective to supply power to each device controller from the remaining separate power supplies in the event of the failure of one of the power supplies connected to any device controller whereby the power supply operatively associated with a processor module can be powered down so that on-line maintenance can be performed in a power-off condition on that processor module while the rest of the multiprocessor system is on-line and functional.

4. A multiprocessor system constructed for continued system operation in the event of a failure of any single com-ponent in the system and comprising, a plurality of separate processor modules, each module including a central processing unit and a local memory, interprocessor bus means including a plurality of interprocessor buses for signaling and data transfer between separate processor modules, at least two interprocessor buses being connected to each processor module, each interprocessor bus being separate and distinct from an input/output system and from a memory bus, a separate bus controller means for each interprocessor bus for determining the priority of data transfers over the interprocessor bus and for controlling transmissions over the interprocessor bus, a plurality of device controllers each adapted for connection to at least one peripheral device, and a plurality of input/output buses each connected be-tween a related processor module and one or more device controllers, each device controller being accessible to at least two processor modules by means of associated input/output buses each of said processor modules including, reading means for reading data from the associated local memory of that processor module for transmission to an interprocessor bus, and writing means for writing data received from an interprocessor bus into a specified location in the associated local memory of that processor module, wherein said reading and writing means are effective to transfer data from the memory of a sender processor module to the memory of a receiver processor module, and wherein each processor module includes a buffer for each other processor module and also includes location pointing means for directing incoming data from an interprocessor bus to a specified location in a related buffer in the memory of a receiver processor module.

5. A method of operating a plurality of separate processor modules concurrently and cooperatively for system control and application processing in a multiprocessor system of the kind in which each processor module includes a central processing unit and a memory for that module, said method comprising connecting each processor module in the system with each other processor module in the system for direct processor module to processor module signalling and data transfer by an interprocessor bus which is separate and distinct from an input/output system and from a memory system bus, connecting each processing module to the interprocessor bus for signalling and data transfer with another module by an interprocessor control in the processor module, determining the priority of data transfers between any two processor modules over the interprocessor bus and controlling transmissions over the interprocessor bus by a bus controller operatively associated with the interprocessor bus and with each interprocessor control of each processor module, transferring data between a processor module and one or more device controllers connected to the processor module by an input/output channel in each processor module, connecting a plurality of peripheral devices to each device controller and controlling the transfer of data between the processor modules and the peripheral devices by the device controllers, connecting each device controller for access by a plurality of different processor modules by a plurality of ports in each device controller and a plurality of input/output busses connected to the respective plurality of ports in each device controller, operatively connecting the input/output channels, input/output busses and device controllers to provide, directly or in combination with the interprocessor bus and a processor module, a path for data transfer between each processor module and any device controller so that a process running in any processor module can access any input/output resource in the system, whether or not the input/output resource is controlled by or attached to the processor module in which the process is running, by the process making requests comprising data trans-missions over the interprocessor bus to a cooperating process in a processor module which controls that input/output resource.