Implementing coherent accelerator function isolation for virtualization
A method, system and computer program product are provided for implementing coherent accelerator function isolation for virtualization in an input/output (IO) adapter in a computer system. A coherent accelerator provides accelerator function units (AFUs), each AFU is adapted to operate independently of the other AFUs to perform a computing task that can be implemented within application software on a processor. The AFU has access to system memory bound to the application software and is adapted to make copies of that memory within AFU memory-cache in the AFU. As part of this memory coherency domain, each of the AFU memory-cache and processor memory-cache is adapted to be aware of changes to data commonly in either cache as well as data changed in memory of which the respective cache contains a copy.
1. A system for implementing coherent accelerator function isolation for virtualization in an input/output (TO) adapter comprising:
a processor;
a hypervisor managing functions associated with the hardware I/O adapter;
the I/O adapter comprising a coherent accelerator including an interface services layer providing Partitionable Endpoint (PE) functions and multiple accelerator function units (AFUs);
each AFU being enabled to operate independently of the other AFUs to perform a computing task that can be implemented within application software on a processor; each AFU having access to system memory bound to the application software and each AFU being adapted to make copies of memory within AFU memory-cache in the AFU; and each of the AFU memory-cache and processor memory-cache being adapted to be aware of changes to data commonly in either cache as well as data changed in memory of which the respective cache contains a copy for memory coherency domain;
the system further comprising a Peripheral Component Interconnect Express (PCIE) host bridge (PHB) providing coherent accelerator PE (Partitionable Endpoint) support including an adapter PE and an AFU PE.
2. The system as recited in claim 1 , wherein said adapter PE communicates with said hypervisor, said interface services layer and said multiple accelerator function units (AFUs) of the I/O adapter; and said AFU PE communicates with each host operating system (OS).
3. The system as recited in claim 1 , wherein said PCI host bridge (PHB) includes a plurality of AFU PEs, and wherein said adapter PE communicates with said hypervisor, and said interface services layer of the I/O adapter; and each of said plurality of AFU PEs communicates with one or more host operating systems (OSs) and a respective one of said multiple accelerator function units (AFUs).
4. The system as recited in claim 1 , wherein said processor and said coherent accelerator communicate changes to individual memory regions represented as cache lines.
5. The system as recited in claim 1 , wherein said hypervisor includes a hypervisor adapter driver which associates each AFU with PHB isolation facilities in support of a PCIE interface.
6. The system as recited in claim 5 , wherein said AFUs are recognized and operated by an operating system (OS) as memory-mapped AFU devices.
7. The system as recited in claim 5 , wherein said hypervisor detects and recovers error involving the interface services layer or AFUs, without requiring the termination of any one operating system (OS) to restore operation of its respective AFU, with the AFUs sharing a common PCI-Express Services Layer (PSL) endpoint function.
8. A system for implementing coherent accelerator function isolation for virtualization in an input/output (IO) adapter comprising:
a processor;
a hypervisor managing functions associated with the hardware I/O adapter;
the I/O adapter comprising a coherent accelerator including an interface services layer providing Partitionable Endpoint (PE) functions and multiple accelerator function units (AFUs);
each AFU being enabled to operate independently of the other AFUs to perform a computing task that can be implemented within application software on a processor; each AFU having access to system memory bound to the application software and each AFU being adapted to make copies of memory within AFU memory-cache in the AFU; and each of the AFU memory-cache and processor memory-cache being adapted to be aware of changes to data commonly in either cache as well as data changed in memory of which the respective cache contains a copy for memory coherency domain;
wherein the system further includes a PCI-Express host bridge (PHB) providing coherent accelerator PE (Partitionable Endpoint) support including an adapter PE and an AFU PE, and wherein said processor and said coherent accelerator use PCI-Express memory read/write operations and each said AFU is associated with a PCI-Express requester ID (RID) for identifying said AFU during the PCI-Express memory read/write operations.