US 2007/0186054 A1Published application

Distributed Cache Coherence at Scalable Requestor Filter Pipes that Accumulate Invalidation Acknowledgements from other Requestor Filter Pipes Using Ordering Messages from Central Snoop Tag

AssigneeAzul Systems, Inc.

InventorsDavid Kruckemyer, Kevin Normoyle, Robert Hathaway

Life of this patent

Estimating the term…

Loading patent…

Abstract

A multi-processor, multi-cache system has filter pipes that store entries for request messages sent to a central coherency controller. The central coherency controller orders requests from filter pipes using coherency rules but does not track completion of invalidations. The central coherency controller reads snoop tags to identify sharing caches having a copy of a requested cache line. The central coherency controller sends an ordering message to the requesting filter pipe. The ordering message has an invalidate count indicating the number of sharing caches. Each sharing cache receives an invalidation message from the central coherency controller, invalidates its copy of the cache line, and sends an invalidation acknowledgement message to the requesting filter pipe. The requesting filter pipe decrements the invalidate count until all sharing caches have acknowledged invalidation. All ordering, data, and invalidation acknowledgement messages must be received by the requesting filter pipe before loading the data into its cache.

Claims20

Dependent claims

A distributed coherency processor comprising:

a plurality of filter pipes for tracking memory requests from a plurality of caches, a requesting filter pipe in the plurality of filter pipes storing a memory-request entry having a request address for a requested cache line;
a central coherency controller that receives memory requests from the plurality of filter pipes, the central coherency controller generating ordering messages and invalidate messages in response to the memory requests;
a snoop tag directory storing snoop entries that indicate sharing caches having a copy of the requested cache line at the request address;
the central coherency controller further for searching the snoop tag directory using the request address from a memory request from a requesting cache, and for sending invalidate messages to sharing filter pipes for the sharing caches identified by the snoop tag directory;
an ordering message sent from the central coherency controller to the requesting filter pipe, the ordering message indicating an order for processing memory requests, the order determined by the central coherency controller;
an invalidate count in the ordering message, the invalidate count indicating a number of the sharing caches in the plurality of caches for the sharing caches identified by the snoop tag directory;
and
a plurality of invalidate acknowledgement messages, generated by the sharing caches in response to the invalidate messages from the central coherency controller, each of the invalidate acknowledgement messages verifying invalidation of the copy of the requested cache line by the sharing cache;
wherein the requesting filter pipe receives the plurality of invalidate acknowledgement messages and releases data in the requested cache line for processing after a number of the plurality of invalidate acknowledgement messages received by the requesting filter pipe matches the invalidate count from the ordering message,
whereby coherency order is determined by the central coherency controller and coherency tracking is performed by the requesting filter pipe in the plurality of filter pipes.

7 more elements

A method for distributed coherency processing comprising:

allocating a memory-request tracking entry in a requesting filter pipe that filters coherency operations requested by a local cache;
generating a request message with a request address for a requested cache line for the local cache and sending the request to a central coherency controller;
receiving a plurality of request messages from a plurality of local caches at the central coherency controller and selecting the request message for processing in a coherency-processing order determined by the central coherency controller to preserve coherency according to coherency rules;
searching a snoop tag directory with the request address from the request message to locate a snoop entry for the requested cache line, the snoop entry indicating sharing caches having a copy of the requested cache line;
when the request message is a request for writing the requested cache line, generating invalidation messages and sending the invalidation messages from the central coherency controller to sharing filter pipes for the sharing caches indicated by the snoop tag directory and generating an invalidate count related to a number of the sharing filter pipes;
generating an ordering message that contains the invalidate count and sending the ordering message from the central coherency controller to the requesting filter pipe;
invalidating copies of the requested cache line at sharing filter pipes in response to the sharing filter pipes receiving the invalidation messages from the central coherency controller and generating invalidation acknowledgement messages;
sending the invalidation acknowledgement messages from the sharing filter pipes to the requesting filter pipe, the invalidation acknowledgement messages bypassing the central coherency controller;
and
releasing the requested cache line for use in the local cache after the requesting filter pipe receives:
(1) the ordering message from the central coherency controller;
(2) a required number of the invalidation acknowledgement messages from the sharing filter pipes, the required number determined by the invalidate count from the ordering message received from the central coherency controller;
and
(3) a data reply containing data for the requested cache line when data is requested;
whereby tracking invalidations by sharing caches is performed by filter pipes for local caches rather than being performed by the central coherency controller which orders operations using the ordering messages.

12 more elements

A cache-coherency system comprising:

coherency controller means for ordering memory requests from a plurality of caches;
a plurality of filter pipe means for storing memory-request tracking entries for memory requests from the plurality of caches;
requesting filter pipe means, in the plurality of filter pipe means, for storing a memory-request tracking entry having a request address for a requested cache line in a requesting cache in the plurality of caches;
request generating means for generating a request message with the request address for the requested cache line for the requesting cache;
request channel means, coupled to the request generating means, for carrying the request message from the requesting filter pipe means to the coherency controller means, the request channel means also for carrying a plurality of request messages from the plurality of filter pipe means for the plurality of caches;
ordering means, in the coherency controller means, for receiving the plurality of request messages from the request channel means, and for selecting the request message for processing in a coherency-processing order determined by the coherency controller means to preserve coherency according to coherency rules;
snoop tag directory means for storing a plurality of snoop entries;
search means for searching the snoop tag directory means with the request address from the request message to locate a snoop entry for the requested cache line, the snoop entry indicating sharing caches in the plurality of caches having a copy of the requested cache line;
invalidation message generation means, activated when the request message is a request for writing the requested cache line, for generating invalidation messages to sharing filter pipes in the plurality of filter pipe means, the sharing filter pipes coupled to sharing caches in the plurality of caches, the sharing caches indicated by the snoop entry;
invalidate count generation means for generating an invalidate count related to a number of the sharing filter pipes;
coherence channel means for carrying the invalidation messages from the coherency controller means to the sharing filter pipes;
ordering message generating means, coupled to the invalidate count generation means and to the ordering means, for generating an ordering message that contains the invalidate count, the ordering message being sent by the coherence channel means from the coherency controller means to the requesting filter pipe means;
sharer invalidation means, in the sharing caches, for invalidating copies of the requested cache line at sharing caches in response to the sharing filter pipes receiving the invalidation messages from the coherency controller means;
sharer message generation means, in the sharing caches, for generating invalidation acknowledgement messages;
reply channel means for carrying the invalidation acknowledgement messages from the sharing caches to the requesting filter pipe means, wherein the invalidation acknowledgement messages bypass the coherency controller means;
and
release means, in the requesting filter pipe means, for releasing the requested cache line for use in the requesting cache after the requesting filter pipe means receives:
(1) the ordering message from the coherency channel means;
(2) a required number of the invalidation acknowledgement messages from the reply channel means from the sharing caches, the required number determined by the invalidate count from the ordering message;
and
(3) a data reply from the reply channel means, the data reply containing data for the requested cache line when data is requested;
whereby tracking invalidations by sharing caches is performed by filter pipe means for caches rather than being performed by the coherency controller means which orders operations using the ordering messages.

19 more elements

Dependencies are inferred from the claim text.