Bull Hn Information Systems, Inc.

Wayne R Buzby

Charles P Ryan

William A Shelly

Sidney L Andress

David A Egolf

Ronald W Yoder

Russell Wayne Guenthner

John E Wilhite

Lowell D McCulley

Robert J Baryla

A multiprocessor write-into-cache data processing system includes a feature for preventing hogging of ownership of a first gateword stored in the memory which governs access to a first common code/data set shared by processes running in the processors by imposing first delays on all other processors in the system while, at the same time, mitigating any adverse effect on performance of processors attempting to access a gateword other than the first gateword. This is achieved by starting a second delay in any processor which is seeking ownership of a gateword other than the first gateword and truncating the first delay in all such processors by subtracting the elapsed time indicated by the second delay from the elapsed time indicated by the first delay.

Multiprocessor write-into-cache system incorporating efficient access to a plurality of gatewords

In a multiprocessor write-into-cache data processing system including: a memory; at least first and second shared caches; a system bus coupling the memory and the shared caches; at least one processor having a private cache coupled, respectively, to each shared cache; method and apparatus for preventing hogging of ownership of a gateword stored in the memory which governs access to common code/data shared by processes running in the processors by which a read copy of the gateword is obtained by a given processor by performing successive swap operations between the memory and the given processor's shared cache, and the given processor's shared cache and private cache. If the gateword is found to be OPEN, it is CLOSEd by the given processor, and successive swap operations are performed between the given processor's private cache and shared cache and shared cache and memory to write the gateword CLOSEd in memory such that the given processor obtains exclusive access to the governed common code/data. When the given processor completes use of the common code/data, it writes the gateword OPEN in its private cache, and successive swap operations are performed between the given processor's private cache and shared cache and shared cache and memory to write the gateword OPEN in memory.

Equal access to prevent gateword dominance in a multiprocessor write-into-cache environment

In a multiprocessor data processing system including: a main memory; at least first and second shared caches; a system bus coupling the main memory and the first and second shared caches; at least four processors having respective private caches with the first and second private caches being coupled to the first shared cache and to one another via a first internal bus, and the third and fourth private caches being coupled to the second shared cache and to one another via a second internal bus; method and apparatus for preventing hogging of ownership of a gateword stored in the main memory and which governs access to common code/data shared by processes running in at least three of the processors. Each processor includes a gate control flag. A gateword CLOSE command, establishes ownership of the gateword in one processor and prevents other processors from accessing the code/data guarded until the one processor has completed its use. A gateword OPEN command then broadcasts a gateword interrupt to set the flag in each processor, delays long enough to ensure that the flags have all been set, writes an OPEN value into the gateword and flushes the gateword to main memory. A gateword access command executed by a requesting processor checks its gate control flag, and if set, starts a fixed time delay after which normal execution continues.

Balanced access to prevent gateword dominance in a multiprocessor write-into-cache environment

When a fault-on-fault condition arises in a data processing system which follows a backup fault procedure in the fault handling process, control is passed to dedicated firmware. Fault flags are reset and information vital to maintaining operating system control is sent to a reserved memory (which can be written to in limited circumstances) under firmware control. Control is then transferred to an Intercept process resident in the reserved memory which attempts to build a stable environment for the operating system to dump the system memory. If possible, a dump is taken, and a normal operating system restart is carried out. If not possible, a message with the vital fault information is issued, and a full manual restart must be taken. Even in the latter case, the fault information is available to help in determining the cause of the fault-on-fault.

Preserving dump capability after a fault-on-fault or related type failure in a fault tolerant computer system

In a multiprocessor data processing system including: a memory, first and second shared caches, a system bus coupling the memory and the shared caches, first, second, third and fourth processors having, respectively, first, second, third and fourth private caches with the first and second private caches being coupled to the first shared cache, and the third and fourth private caches being coupled to the second shared cache, gateword hogging is prevented by providing a gate control flag in each processor. Priority is established for a processor to next acquire ownership of the gate control word by: broadcasting a &ldquo;set gate control flag&rdquo; command to all processors such that setting the gate control flags establishes delays during which ownership of the gate control word will not be requested by another processor for predetermined periods established in each processor. Optionally, the processor so acquiring ownership broadcasts a &ldquo;reset gate control flag&rdquo; command to all processors when it has acquired ownership of the gate control word.

Gateword acquisition in a multiprocessor write-into-cache environment

A fault number is utilized by microcode fault handling to index into a fault array pointer table containing a plurality of pointers to entry descriptors describing fault handling routines. The pointer resulting from the indexing is utilized to retrieve an entry descriptor. The entry descriptor is verified and if valid, is utilized to setup the environment for the appropriate fault handling routine and to enter such. The fault array pointer table is located in a reserved memory that cannot be overwritten by I/O. During the boot process, the fault array pointer table entries, along with a fault-on-fault pointer are updated to point at entry descriptors stored in the reserved memory. Additionally, the fault-on-fault entry descriptor that rebuilds the processor environment if necessary from information in reserved memory.

Fault handling in a data processing system utilizing a fault vector pointer table

Fault vector pointer table

In a NUMA architecture, processors in the same CPU module with a processor opening a spin gate tend to have preferential access to a spin gate in memory when attempting to close the spin gate. This &ldquo;unfair&rdquo; memory access to the desired spin gate can result in starvation of processors from other CPU modules. This problem is solved by &ldquo;balking&rdquo; or delaying a specified period of time before attempting to close a spin gate whenever either one of the processors in the same CPU module just opened the desired spin gate, or when a processor in another CPU module is spinning trying to close the spin gate. Each processor detects when it is spinning on a spin gate. It then transmits that information to the processors in other CPU modules, allowing them to balk when opening spin gates.

Gate close balking for fair gating in a nonuniform memory architecture data processing system

Gate close failure notification for fair gating in a nonuniform memory architecture data processing system

In order to gather, store temporarily and efficiently deliver safestore information in a CPU having data manipulation circuitry including a register bank, first and second serially oriented safestore buffers are employed. At suitable times during the processing of information, a copy of the instantaneous contents of the register bank is transferred into the first safestore buffer. After a brief delay, a copy of the first safestore buffer is transferred into the second safestore buffer. If a call for a domain change (which might include a process change or a fault) is sensed, a safestore frame is sent to cache, and the first safestore buffer is loaded from he second safestore buffer rather than from the register bank. Later, during a climb operation, if a restart of the interrupted process is undertaken and the restoration of the register bank is directed to be taken from the first safestore buffer, this source, rather than the safestore frame stored in cache, is employed to obtain a corresponding increase in the rate of restart. In one embodiment, the transfer of information between the register bank and the safestore buffers is carried out on a bit-by-bit basis to achieve additional flexibility of operation and also to conserve integrated circuit space.

Fast domain switch and error recovery in a secure CPU architecture

In order to gather, store temporarily and efficiently deliver (if needed) safestore information in a fault tolerant central processing unit having data manipulation circuitry including a plurality of software visible registers, a safestore memory for storing the contents of the plurality of software visible registers, after a data manipulation operation, is provided. Iterative execution instructions subject to a page fault are specially handled in that, during execution, status information indicative of the ongoing status and valid intermediate results are additionally stored in the safestore memory. Then, in the event of a page fault encountered during the execution of the iterative execution instruction, execution is suspended until access to a valid copy of the missing page is obtained. When a valid copy becomes available, the execution of the iterative execution instruction is restarted at the point at which the valid intermediate results had been obtained prior to occurrence of the page fault.

Growing community of inventors

Phoenix, AZ, United States of America

Wayne R Buzby