International Business Machines Corporation

Other

George William Daly, Jr

James Stephen Fields, Jr

Bartholomew Blaner

Joseph Gerald McDonald

Ross Boyd Leavens

Paul Umbarger

Ravi Kumar Arimilli

Michael Steven Siegel

Brian Mitchell Bass

Jeffrey Haskell Derby

Guy L Guthrie

William John Starke

Jeff A Stuecheli

Donald G Grice

David W Cummings

Appoloniel N Tankeh

Derek E Williams

Jeffrey A Stuecheli

Warren Edward Maule

Kenneth Lee Wright

Richard Loise Arndt

Thomas K Heller, Jr

Thomas J Heller

A computer system having a plurality of processing resources, including a sub-system for scheduling and dispatching processing jobs to a plurality of hardware accelerators, the subsystem further comprising a job requestor, for requesting jobs having bounded and varying latencies to be executed on the hardware accelerators; a queue controller to manage processing job requests directed to a plurality of hardware accelerators; and multiple hardware queues for dispatching jobs to the plurality of hardware acceleration engines, each queue having a dedicated head of queue entry, dynamically sharing a pool of queue entries, having configurable queue depth limits, and means for removing one or more jobs across all queues.

Dynamically configurable hardware queues for dispatching jobs to a plurality of hardware acceleration engines

A coherent attached processor proxy (CAPP) that participates in coherence communication in a primary coherent system on behalf of an external attached processor maintains, in each of a plurality of entries of a CAPP directory, information regarding a respective associated cache line of data from the primary coherent system cached by the attached processor. In response to initiation of recovery operations, the CAPP transmits, in a generally sequential order with respect to the CAPP directory, multiple memory access requests indicating an error for addresses indicated by the plurality of entries. In response to a snooped memory access request that targets a particular address hitting in the CAPP directory during the transmitting, the CAPP performs a coherence recovery operation for the particular address prior to a time indicated by the generally sequential order.

Accelerated recovery for snooped addresses in a coherent attached processor proxy

A technique for maintaining input/output (I/O) command ordering on a bus includes assigning a channel identifier to I/O commands of an I/O stream. In this case, the channel identifier indicates the I/O commands belong to the I/O stream. A command location indicator is assigned to each of the I/O commands. The command location indicator provides an indication of which one of the I/O commands is a start command in the I/O stream and which of the I/O commands are continue commands in the I/O stream. The I/O commands are issued in a desired completion order. When a first one of the I/O commands does not complete successfully, the I/O commands in the I/O stream are reissued on the bus starting at the first one of the I/O commands that did not complete successfully.

Techniques for write-after-write ordering in a coherency managed processor system that employs a command pipeline

A method and apparatus for invalidating cache lines during direct memory access (DMA) write operations are disclosed. Initially, a multi-cache line DMA request is issued by a peripheral device. The multi-cache line DMA request is snooped by a cache memory. A determination is then made as to whether or not the cache memory includes a copy of data stored in the system memory locations to which the multi-cache line DMA request are directed. In response to a determination that the cache memory includes a copy of data stored in the system memory locations to which the multi-cache line DMA request are directed, multiple cache lines within the cache memory are consecutively invalidated.

Method and apparatus for invalidating cache lines during direct memory access (DMA) write operations

A data processing system includes at least a first processing node having an input/output (I/O) controller and a second processing including a memory controller for a memory. The memory controller receives, in order, pipelined first and second DMA write operations from the I/O controller, where the first and second DMA write operations target first and second addresses, respectively. In response to the second DMA write operation, the memory controller establishes a state of a domain indicator associated with the second address to indicate an operation scope including the first processing node. In response to the memory controller receiving a data access request specifying the second address and having a scope excluding the first processing node, the memory controller forces the data access request to be reissued with a scope including the first processing node based upon the state of the domain indicator associated with the second address.

Data processing system and method that permit pipelining of I/O write operations and multiple operation scopes

A method of interfacing two components of a computing system is provided wherein the method includes providing a pair of unidirectional, point-to-point buses to transmit data between a master bus controller of the computing system and a slave bus controller of a processor unit of the computing system. The method also includes providing means for transmitting a command packet with an address associated with data pertaining to the command from the master bus controller to the slave bus controller. In addition, the method includes providing means for determining by the slave bus controller whether the slave bus controller can accept the command. The method further includes providing means for transmitting an acknowledgement from the slave bus controller to the master bus controller after the slave bus controller receives a first signaling interval for the command packet if the slave bus controller can accept the command packet.

Method and system for interfacing components of a computing system with a pair of unidirectional, point-to-point buses

Method for interfacing components of a computing system with a pair of unidirectional, point-to-point buses

A method and apparatus for invalidating entries within a translation control entry (TCE) cache are disclosed. A host bridge is coupled between a group of processors and a group of adaptors. The host bridge includes a TCE cache. The TCE cache contains the most-recently use copies of TCEs in a TCE table located in a system memory. In response to a modification to a TCE in the TCE table by one of the processors, a memory mapped input/output (MMIO) Store is sent to a TCE invalidate register to specify an address of the modified TCE. The data within the TCE invalidate register is then utilized to generate a command for invalidating an entry in the TCE cache containing an unmodified copy of the modified TCE in the TCE table. The command is subsequently sent to the host bridge to invalidate the entry in the TCE cache.

Method and apparatus for invalidating entries within a translation control entry (TCE) cache

A method, system and computer program product for handling write requests in a data processing system is disclosed. The method comprises receiving on an interconnect bus a first write request targeted to a first address and receiving on the interconnect bus a subsequent second write request targeted to a subsequent second address. The subsequent second write request is completed prior to completing the first write request, and, responsive to receiving a read request targeting the second address before the first write request has completed, data associated with the second address of the second write request is supplied only after the first write request completes.

Method to preserve ordering of read and write operations in a DMA system by delaying read access

A data processing system that provides a DMA Exclusive state that enables pipelining of Input/Output (I/O) DMA Write transactions. The data processing system includes a system processor, a system bus, a memory, a plurality of I/O components and an I/O processor. The data processing system further comprises operational protocol providing a pair of instructions/commands that are utilized to complete a DMA Write operation. The pair of instructions is DMA_Write_No_Data and DMA_Write With_Data. DMA_Write_No_Data is an address-only operation on the system bus that is utilized to acquire &ldquo;DMA ownership&rdquo; of a cache line that is to be written. The initial ownership of the cache line is marked by a weak DMA state (D1), which indicates that the cache line is being held for writing to the memory, but that the cache line cannot yet force a retry of snooped operations. When each preceding DMA Write operation has completed or each corresponding DMA_Write_No_Data operation has been placed in a DMA Exclusive (D2) state, then the next cache line in a weak DMA state transitions to a DMA Exclusive (D2) state, which forces a retry of snooped operations until the DMA_Write_With_Data transaction to memory is completed. Accordingly, DMA_Write_No_Data operations that are provided sequentially may be completed in a parallel manner on the system bus although the corresponding DMA_Write_With_Data is held until a DMA Exclusive state attaches to the cache line. Also, the DMA_Write_With_Data may be completed out of order and once the write of the cache line is completed, the coherency state transitions from the DMA Exclusive state to the MESI Exclusive (E) state or the MESI Invalidate (I) state depending on processor operating characteristics.

DMA exclusive cache state providing a fully pipelined input/output DMA write mechanism

A method and data processing system that supports pipelining of Input/Output (I/O) DMA Write transactions. An I/O processor's operational protocol is provided with a pair of instructions/commands that are utilized to complete a DMA Write operation. The instructions are DMA_Write_No_Data and DMA_Write_With_Data. DMA_Write_No_Data is an address-only operation on the system bus that is utilized to acquire ownership of a cache line that is to be written. The ownership of the cache line is marked by a weak DMA state, which indicates that the cache line is being held for writing to the memory, but that the cache line cannot yet force a retry of snooped operations. When each preceding DMA Write operation has completed or each corresponding DMA_Write_No_Data operation has been placed in a DMA Exclusive state, then the weak DMA state is changed to a DMA Exclusive state, which forces a retry of snooped operations until the write transaction to memory is completed. In this way, DMA Writes that are provided sequentially may be issued in a parallel manner on the system bus and their corresponding DMA_Write_No_Data operations may be completed in any order, but cannot be made DMA Exclusive unless the above conditions are satisfied. Further, once a DMA Exclusive state is acquired, a DMA_Write_With_Data may be issued for each of the sequential DMA Write operations in the DMA Exclusive state. The DMA_Write_With_Data may then be completed out-of-order with respect to each other. However, the system processor is sent the completion messages in the sequential order of the DMA Write operations, thus adhering to the processor requirements for ordered operations while providing fully-pipelined (parallel) execution of the DMA transactions.

Growing community of inventors

Austin, TX, United States of America

George William Daly, Jr