6. System Interface Operations

6.21 Support for External Duplicate Tags

Some system designs implement an external duplicate copy of the secondary cache tags to reduce the coherency request latency and also filter out unnecessary external coherency requests made to the R10000 processor.

For such systems, it must be remembered that blocks may reside in either the secondary cache or in the outgoing buffer. During the address cycle of processor block read requests, the secondary cache block former state is provided. The external agent may use this information to maintain the external duplicate tags.

Typically, in a multiprocessor system using the cluster bus, the cluster coordinator specifies a free request number for an external coherency request. However, in a system using a duplicate-tag or directory-based coherency protocol, where the CohPrcReqTar mode bit is negated, the cluster coordinator may specify a busy request number for an external coherency request, providing each targeted R10000 processor has the request number busy due to an outstanding processor coherency request from another processor.

For example, suppose the processor in master state issues a processor coherent block read or upgrade request. The processors in slave state observe the processor request as an external coherency request that targets the external agent only, causing the associated request number to become busy. The cluster coordinator checks the duplicate tag or directory structure to determine if the block resides in the cache of one of the processors that was in slave state. If necessary, the cluster coordinator issues an external coherency request targeted at one or more of the processors that were in slave state. By using the same request number as the original processor request, this external coherency request does not consume a free request number, and allows a potential processor coherency data response to be supplied as an external block data response to the original processor request.

Generated with CERN WebMaker