Abstract: An effectively “conditional”, cast out operation or cast out portion of a combined operation including a related data access may be cancelled by the combined response to the operation. The combined response logic receives coherency state and/or LRU position information for cache lines corresponding to the cast out victim within snoopers and vertically in-line storage. The combined response logic may also receive information regarding the presence of shared or invalid cache lines in snoopers or lower level storage within the congruence class for the victim, or information regarding the read-once nature of the data access target. Based on these responses, the combined response logic determines whether the cast out should be cancelled and, if so, selects and drives the appropriate combined response code.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: Logically in line caches within a multilevel cache hierarchy are jointly controlled by single cache controller. By combining the cache controller and snoop logic for different levels within the cache hierarchy, separate queues are not required for each level. During a cache access, cache directories are looked up in parallel. Data is retrieved from an upper cache if hit, or from the lower cache if the upper cache misses and the lower cache hits. LRU units may be updated in parallel based on cache directory hits. Alternatively, the lower cache LRU unit may be updated based on cache memory accesses rather than cache directory hits, or the cache hierarchy may be provided with user selectable modes of operation for both LRU unit update schemes. The merged vertical cache controller mechanism does not require the lower cache memory to be inclusive of the upper cache memory. A novel deallocation scheme and update protocol may be implemented in conjunction with the merged vertical cache controller mechanism.

Abstract: A method of avoiding deadlocks in cache coherency protocol for a multi-processor computer system, by loading a memory value into a plurality of cache blocks, assigning a first coherency state having a higher collision priority to only one of the cache blocks, and assigning one or more additional coherency states having lower collision priorities to all of the remaining cache blocks. Different system bus codes can be used to indicate the priority of conflicting requests (e.g., DClaim operations) to modify the memory value. The invention also allows folding or elimination of redundant DClaim operations, and can be applied in a global versus local manner within a multi-processor computer system having processing units grouped into at least two clusters.

Abstract: A modified MESI cache coherency protocol is implemented within a level two (L2) cache accessible to a processor having bifurcated level one (L1) data and instruction caches. The modified MESI protocol includes two substates of the shared state, which denote the same coherency information as the shared state plus additional information regarding the contents/coherency of the subject cache entry. One substate, SIC0, indicates that the cache entry is assumed to contain instructions since the contents were retrieved from system memory as a result of an instruction fetch operation. The second substate, SIC1, indicates the same information plus that a snooped flush operation hit the subject cache entry while it's coherency was in the first shared substate. Once the first substate is entered, the coherency state does not transition to the invalid state unless an operation designed to invalidate instructions is received.

Abstract: A method of avoiding deadlocks in cache coherency protocol for a multi-processor computer system, by loading a memory value into a plurality of cache blocks, assigning a first coherency state having a higher collision priority to only one of the cache blocks, and assigning one or more additional coherency states having lower collision priorities to all of the remaining cache blocks. Different system bus codes can be used to indicate the priority of conflicting requests (e.g., DClaim operations) to modify the memory value. The invention also allows folding or elimination of redundant DClaim operations, and can be applied in a global versus local manner within a multi-processor computer system having processing units grouped into at least two clusters.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (Mu) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the Mu state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The Mu state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the Mu state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (Mu) cache state to indicate that a value held in 5 a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the Mu state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The Mu state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the Mu state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: Upon snooping a combined data access and cast out/deallocate operation initiating by a horizontal storage device, snoop logic determines appropriate responses to both the data access and the cast out/deallocate based upon the presence and coherency state of the target of the data access within a corresponding storage device, the presence and coherency state of the victim of the cast out/deallocate within the corresponding storage device, and the presence of an invalid entry within the corresponding storage device in a congruence class including both the target and the victim. The appropriate responses are “merged”, transmitted together in response to the combined operation as either a single response code or discrete response codes within a single response. The coherency state and LRU position of the selected victim for the cast out/deallocate portion of the combined operation may also be appended to the response to facilitate data storage management.

Abstract: In response to receiving a combined address for related data access and cast out operations, including an index identifying a congruence class containing both the target of the data access and the victim of the cast out, a single directory access is performed utilizing the index to locate the congruence class. Address tags within the congruence class are then compared to the address tag for the data access operation and the address tag for the cast out operation concurrently, generating separate hit signals as appropriate. Only a single directory access is required, however, rather than two separate directory accesses as required in the known art, taking advantage of the fact that both the data access target and the cast out victim belong to a single congruence class. Response latency is also improved, as is address bus bandwidth utilization.

Abstract: A cache coherency protocol uses a “Tagged” coherency state to track responsibility for writing a modified value back to system memory, allowing intervention of the value without immediately writing it back to system memory, thus increasing memory bandwidth. The Tagged state can migrate across the caches (horizontally) when assigned to a cache line that has most recently loaded the modified value. Historical states relating to the Tagged state may further be used. The invention may also be applied to a multi-processor computer system having clustered processing units, such that the Tagged state can be applied to one of the cache lines in each group of caches that support separate processing unit clusters. Priorities are assigned to different cache states, including the Tagged state, for responding to a request to access a corresponding memory block. Any tagged intervention response can be forwarded only to selected caches that could be affected by the intervention response, using cross-bars.

Abstract: Combined response logic for a bus receives a combined data access and cast out/deallocate operation initiating by a storage device within a specific level of a storage hierarchy, with a coherency state and LRU position of the cast out/deallocate victim appended. Snoopers on the bus drive snoop responses to the combined operation with the coherency state and/or LRU position of locally-stored cache lines corresponding to the victim appended. The combined response logic determines, from the coherency state and LRU position information appended to the combined operation and the snoop responses, whether an update of the LRU position and/or coherency state of a cache line corresponding to the victim within one of the snoopers is required. If so, the combined response logic selects a snooper storage device to have at least the LRU position of a respective cache line corresponding to the victim updated, and appends an update command identifying the selected snooper to the combined response.

Abstract: A cache coherency protocol uses a “Tagged” coherency state to track responsibility for writing a modified value back to system memory, allowing intervention of the value without immediately writing it back to system memory, thus increasing memory bandwidth. The Tagged state can migrate across the caches (horizontally) when assigned to a cache line that has most recently loaded the modified value. Historical states relating to the Tagged state may further be used. The invention may also be applied to a multi-processor computer system having clustered processing units, such that the Tagged state can be applied to one of the cache lines in each group of caches that support separate processing unit clusters. Priorities are assigned to different cache states, including the Tagged state, for responding to a request to access a corresponding memory block. Any tagged intervention response can be forwarded only to selected caches that could be affected by the intervention response, using cross-bars.

Abstract: A method of avoiding deadlocks in cache coherency protocol for a multi-processor computer system, by loading a memory value into a plurality of cache blocks, assigning a first coherency state having a higher collision priority to only one of the cache blocks, and assigning one or more additional coherency states having lower collision priorities to all of the remaining cache blocks. Different system bus codes can be used to indicate the priority of conflicting requests (e.g., DClaim operations) to modify the memory value. The invention also allows folding or elimination of redundant DClaim operations, and can be applied in a global versus local manner within a multi-processor computer system having processing units grouped into at least two clusters.

Abstract: A combined address bus transaction contains the address tag for a data access operation target, the address tag for a victim to be replaced, and the address index field identifying the congruence class including both the target and the victim. Directory state information such as coherency state and/or LRU position for the cast out victim may also be appended to the index field and target and victim address tags within the bus operation. Address bus bandwidth utilization is thereby improved, eliminating duplicate transmission of the index field employed by separate data access and cast out operations in accordance with the existing practice. The victim may be prospectively selected concurrently with the determination of whether the target may be found within the storage device forming the combined address, improving overall performance for that device.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: In cancelling the cast out portion of a combined operation including a data access related to the cast out, the combined response logic explicitly directs the storage device initiating the combined operation not to allocate storage for the target of the data access. Instead, the target of the data access may be passed directly to an in-line processor core without storage, may be stored in a horizontal storage device, or may be stored in an in-line, noninclusive, lower level storage device. Cancellation of the cast out thus defers any latency associated with writing the cast out victim to system memory while maximizing utilization of available storage with acceptable tradeoffs in data access latency.

Abstract: A cache coherency protocol uses a “Exclusive-Deallocate” (ED) coherency state to indicate that a particular value is currently held in an upper level cache in an exclusive, unmodified form (not shared with any other caches of the computer system, including caches associated with the same processing unit), so that the value can conveniently be modified without any lower level bus transactions since no lower level caches have allocated a line for the value. If the value is subsequently modified in the upper level cache, its coherency state is simply switched to “modified” without the need for any bus transactions. Conversely, if the value is evicted from the upper level cache without ever having been modified, it can be loaded into the lower level cache with a coherency state indicating that the lower level cache contains the unmodified value exclusive of all other caches in other processing units of the computer system.

Abstract: A cache and method of maintaining cache coherency in a data processing system are described. The data processing system includes a plurality of processors and a plurality of caches coupled to an interconnect. According to the method, a first data item is stored in a first of the caches in association with an address tag indicating an address of the first data item. A coherency indicator in the first cache is set to a first state that indicates that the tag is valid and that the first data item is invalid. Thereafter, the interconnect is snooped to detect a data transfer initiated by another of the plurality of caches, where the data transfer is associated with the address indicated by the address tag and contains a valid second data item. In response to detection of such a data transfer while the coherency indicator is set to the first state, the first data item is replaced by storing the second data item in the first cache in association with the address tag.