International Business Machines Corporation

Other

Motorola Corporation

Thomas Albert Petersen

Jose Melanio Nunez

Alexander Edward Okpisz

James Nolan Hardage

Srinath Audityan

Brian R Konigsburg

Bruce Joseph Ronchetti

Ronny Lee Arnold

Scott Ives Remington

Marie Jeannette Sullivan

Amy May Tuvell

Robert Charles Podnar

Joseph Yih Chang

Where a null response can be expected from devices snooping a load operation, data may be used by a requesting processor prior to the coherency response window. A null snoop response may be determined, for example, from the availability of the data without a bus transaction. The capability of accelerating data in this fashion requires only a few simple changes in processor state transitions, required to permit entry of the data completion wait state prior to the response wait state. Processors may forward accelerated data to execution units with the expectation that a null snoop response will be received during the coherency response window. If a non-null snoop response is received, an error condition is asserted. Data acceleration of the type described allows critical data to get back to the processor without waiting for the coherency response window.

Data received before coherency window for a snoopy bus

A system and method are provided for maintaining cache coherency in symmetric multiprocessor system by having logic for performing snoop queries separate from logic for performing snoop actions. This logic split permits increased throughput because pending snoop queries must no longer wait until snoop actions complete. Two state machines are disclosed: a pipeline state machine for snoop queries; and a snoop action state machine, for executing the split logic handling snoop requests.

Optimizing pipelined snoop processing

A method and system of facilitating storage accesses within a multiprocessor system subsequent to a synchronization instruction by a local processor consists of determining if data for the storage accesses is cacheable and if there is a &ldquo;hit&rdquo; in a cache. If both conditions are met, the storage accesses return the data to the local processor. The storage accesses have an entry on an interrupt table which is used to discard the returned data if a snoop kills the line before the synchronization instruction completes. After the cache returns data, a return data bit is set in the interrupt table. A snoop killing the line sets a snooped bit in the interrupt table. Upon completion of the synchronization instruction, any entries in the interrupt table subsequent to the synchronization instruction that have the return data bit and snooped bit set are flushed. The flush occurs because the data returned to the local processor due to a &ldquo;cacheable hit&rdquo; subsequent to the synchronization instruction was out of order with the snoop and the processor must flush the data and go back out to the system bus for the new data.

Method and system for speculatively processing a load instruction before completion of a preceding synchronization instruction

A multiprocessor computer system including a set of processors where each processor in the set includes an execution unit for issuing operations and a processor queue suitable for queuing previously issued and still pending operations. The multiprocessor further includes means for forwarding operations issued by the processor to the processor queue and to an operation block queue of a memory subsystem that is connected to the multiprocessor. The depth of (i.e., the number of entries in) the operation block queue matches the depth of the processor queue. The processor queue, when full, inhibits the processor from issuing additional operations. In this manner, an operation issued by the processor is guaranteed an available entry in the operation block queue of the memory subsystem thereby eliminating the need for operation retry circuitry and protocols such as handshaking. Preferably, each processor queue includes a processor load queue and a processor store queue and the operation block queue includes a load queue and a store queue. In this embodiment, the depth of each of the processor load and store queues matches the depth of the operation block load and store queues respectively. In the preferred embodiment, the operation block is comprised of a load miss block that includes the operation block load queue and a store miss block that includes the operation block store queue. Still further preferably, the operation block store queue includes a set of store queues corresponding to the set of processors and the operation block load queue includes a set of load queues corresponding to the set of processors. Each queue entry preferably includes state information indicative of the status of the corresponding entry.

Queue resource tracking in a multiprocessor system

A multi processor computer system including a set of processors connected to a memory subsystem via a local interconnect. The memory subsystem includes a load miss block suitable for queuing a first processor load operation that misses in an L cache of the first processor and a store miss block suitable for queuing store type operations. The subsystem further includes an arbiter suitable for receiving queued operations from the load and store miss blocks. The arbiter is further configured for selecting one of the received operations and initiating the selected operation. The subsystem further includes means for snooping the address associated with the first processor load operation when the first processor load operation is selected and initiated by the arbiter. The subsystem further includes a snoop control block adapted to receive a snoop response from a second processor associated with the memory subsystem. The snoop control block is further adapted to queue a store type operation in the store miss block if the snoop response from the second processor is modified. The subsystem is configured to link the store type operation with the first load operation when the store type operation is initiated. When the linked operations complete (together), the data associated with the store type operation, which is preferably written to an L or lower level cache, will also satisfy the first load operation. The local interconnect is preferably comprised of a unidirectional bus. In the preferred embodiment, the load and store blocks each include control pipelines with corresponding stages wherein each stage has its own validity information. In this embodiment the corresponding stages of the load miss and store miss blocks are linked by simultaneously validating a first stage of the load miss block when the forwarding operation is initiated (i.e., when the forwarding operation wins arbitration by the arbiter).

Local cache-to-cache transfers in a multiprocessor system

A method and system for performing write-through store operations of valid data of varying sizes in a data processing system, where the data processing system includes multiple processors that are coupled to an interconnect through a memory hierarchy, where the memory hierarchy includes multiple levels of cache, where at least one lower level of cache of the multiple of levels of cache requires store operations of all valid data of at least a predetermined size. First, it is determined whether or not a write-through store operation is a cache hit in a higher level of cache of the multiple levels of cache. In response to a determination that cache hit has occurred in the higher level of cache, the write-through store operation is merged with data read from the higher level of cache to provide a merged write-through operation of all valid data of at least the predetermined size to a lower level of cache. The merged write-through operation is performed in the lower level of cache, such that write-through operations of varying sizes to a lower level of cache which requires write operations of all valid data of at least a predetermined size are performed with data merged from a higher level of cache.

Method and system for write-through stores of varying sizes

The present invention provides a method and apparatus for allowing autonomous, way specific tag updates. More specifically, the invention provides way specific tag and status updates while concurrently allowing reads of the ways not currently being updated. If a read hit is determined, then the read is processed in a typical fashion. However, if the read is a read miss and one of the ways is flagged as being updated, then all ways are read again once the specific way has completed its updated.

Autonomous way specific tag update

A multiprocessor computer system including a set of processors connected to a memory subsystem via a local interconnect. The memory subsystem includes a load miss block suitable for queuing a first processor load operation that misses in an L1 cache of the first processor and a store miss block suitable for queuing store type operations. The subsystem further includes an arbiter suitable for receiving queued operations from the load and store miss blocks. The arbiter is further configured for selecting one of the received operations and initiating the selected operation. The subsystem further includes means for snooping the address associated with the first processor load operation when the first processor load operation is selected and initiated by the arbiter. The subsystem further includes a snoop control block adapted to receive a snoop response from a second processor associated with the memory subsystem. The snoop control block is further adapted to queue a store type operation in the store miss block if the snoop response from the second processor is modified. The subsystem is configured to link the store type operation with the first load operation when the store type operation is initiated. When the linked operations complete (together), the data associated with the store type operation, which is preferably written to an L2 or lower level cache, will also satisfy the first load operation. The local interconnect is preferably comprised of a unidirectional bus.

Data source arbitration in a multiprocessor system

Data loaded from system memory to a cache within a multiprocessor system is set to the exclusive coherency state if no other cache or processor has a copy of that data. Subsequent accesses to the data by another processor or cache which are snooped by the data owner result in an exclusive intervention by the data owner. The data owner sources the data to and shares the data with the requesting device on a read and transfers exclusive ownership of the data to the requesting device on a read with intent to modify. Unmodified intervention with cache-to-cache transfers over possibly much slower accesses to memory is thus supported by the multiprocessor system without requiring additional tag or status bits in the cache directories, saving a significant area.

6XX bus with exclusive intervention

A queuing apparatus associated with a processor includes at least one static queue (,), an index generator (,), at least one index queue (,), and a static queue accessing arrangement. Each static queue (,) has a number of static queue locations (,), each for storing a static queue entry and an availability indicator (,) for indicating an availability status of the respective static queue location. The index generator (,) uses information from the static queue (,) to provide a unique index value for each static queue entry, the index value for a particular static queue entry identifying the static queue location (,) containing the particular static queue entry. Each index queue (,) has a number of index queue locations (,), each for storing one of the index values provided by the index generator (,). The static queue accessing arrangement retrieves a selected index value from a particular index queue location (,), and uses the selected index value to retrieve the static queue entry with which the selected index value is associated. Multiple index queues (,) facilitate prioritization of static queue entries, and reprioritization by transferring index queue values from one index queue to another. The index queues (,) also facilitate compaction of unnecessary static queue entries.

Static queue and index queue for storing values identifying static queue locations

A multiprocessor computer system including a multiprocessor device preferably comprised of a set of processors, each including a respective L,cache. The multiprocessor is preferably fabricated as a single device. The computer system includes a memory subsystem comprised of a load miss block adapted for queuing a load operation issued by a first processor that misses in an L,cache of the first processor and a store miss block adapted for queuing store type operations. An arbiter of the memory subsystem is configured to receive queued operations from the load and store miss blocks and further configured to select and initiate one of the received operations. The subsystem further includes means for forwarding the address associated with the load miss operation to a lower level cache and means for receiving a hit/miss response from the lower level cache. In the preferred embodiment, the load miss block is adapted to detect the response from lower level cache and to request a bus interface unit to fetch data via a system bus if the lower level cache responds with a miss. The bus interface unit is configured to signal the load miss block when a first portion of the fetched data is available. In response thereto, the load miss block is configured to initiate a forwarding operation that returns the first potion of the data to the requesting processor if the forwarding operation can be initiated without displacing a valid load miss operation. The store and load miss block preferably each include separate store miss queues for each processor of the multiprocessor. The bus interface unit is preferably further configured to signal the load miss block when the entire granule (i.e., cache line) of requested data is available. The forwarding operation is preferably initiated if a first stage of a load miss block pipeline is invalid at some point after the first portion data is available, but before the entire requested data is available.

Critical word forwarding in a multiprocessor system

Where a plurality of ordered transactions are received for data transfers on a pipelined bus, each transaction in the series is initiated before all prospective retry responses to the preceding ordered transactions may be asserted. The address responses to all preceding ordered transfers are then monitored in connection with performance of the newly initiated transfer. If a retry response to any preceding ordered transaction is asserted, a self-initiated retry response for all subsequent transactions, including the newly initiated transfer, is also asserted. The system-retried transactions and all succeeding, ordered transactions are immediately reattempted. The overlapping performance of the ordered transfers reduces the latency of non-retried transfers, achieving performance comparable to non-ordered transactions. Even where a retry response is asserted, the total latency required for completion of both transactions in the ordered pair is reduced by at least a portion of the address-to-response latency, so that the impact of ordering requirements on system performance is minimized. Strict ordering is thus enforced while taking full advantage of the pipelined nature of the bus to maximize utilization of the bus bandwidth.

Optimization of ordered stores on a pipelined bus via self-initiated retry

The present invention provides a method and apparatus for optimizing bus utilization while maintaining read and write coherence. More specifically the invention provides bus utilization optimization by prioritizing read transactions before write transactions, where there is no collision pending. When a collision pending is determined, then the read and write transactions are processed according to the age of the transaction(s) allowing for data coherency.

Bus optimization with read/write coherence including ordering responsive to collisions

A method for improving data processing in an L,cache for symmetrical multiprocessing systems consists of efficient execution of cache control instructions without having to give up the data bandwidth provided by a greater byte coherency granule. The L,cache has a coherency granule size within its data array and is divided into a target sector and an alternate sector. Additionally, the coherency granule has a plurality of MESI bits, which define sector write enables, and data write enables. By determining the states of the target sector and/or the alternate sector a series of L,cache control instructions are performed to signal the L,cache to hit. If a hit occurs corresponding data will be either written into or read out of the data array.

Method for supporting cache control instructions within a coherency granule

Storage access blocking instructions, such as the EIEIO instruction implemented within the PowerPC architecture, block other storage access instructions at the bus interface stage as opposed to the execute stage. Therefore, cacheable instructions, and other similar instructions, are allowed to complete without being blocked by such an EIEIO instruction not ordered by the EIEIO instruction.

Growing community of inventors

Austin, TX, United States of America

Thomas Albert Petersen