Using read current transactions for improved performance in directory-based coherent i/o systems

Abstract

A shared memory provides data access to a plurality of agents (e.g., processor, cells of processors, I/O controllers, etc.) and includes a memory and a memory controller. The memory controller selectively provides memory access to the agents in both coherent and read current modes of operation. In the coherent mode, the memory controller ensures that the data stored in system memory is accurately and precisely mirrored in all subservient copies of that data as might typically be stored in agent cache memories. Using, for example, a MESI type protocol, the memory controller limits access to memory so that only an “owner” or a particular portion or line of memory has write access and that, during the extension of these write privileges, no other agent has a valid copy of the data subject to being updated. Thus, the memory controller implements a first set of rule in the coherent mode of operation to insure that all copies of data stored by the agents are coherent with data stored in the memory. In a read current mode of access, a read-once segment of data is copied to an agent with the agent implementing a second set of rules to minimize or eliminate the possibility that the data might become stale prior to use or that it be misused by another agent or process. Thus, in the read current, an “uncontrolled” copy of the data is released subject to certain restrictions to be implemented by the receiving agent as defined by a second set of rules. For example, these rules require that the agent's copy of data be provided as an output and then invalidated within a predetermined period of time, that the agent have limit access to the memory during any set of fetch operations to no more than a predetermined maximum block size. Also included is a requirement that the copy of data include only that data contained within a range of memory addresses, the range beginning within a predetermined block of memory addresses and continuing through an end of block address. These limitations limit the amount of data that might be subject to misuse, particularly in the case of a failure resulting in the inability of a requesting agent to complete a transaction according to the rules.