Universal DMA (direct memory access) architecture
A universal DMA (Direct Memory Access) engine can be dynamically configured to function in either a receive or transmit mode. DMAs are logically assembled and bound as needed, without limitation to a fixed, pre-determined number of receive engines and transmit engines. Because a DMA engine may be dynamically assembled to support the flow of data in either direction, varied usage models are enabled, and components used to assemble a receive DMA engine for one application may be subsequently used to assemble a transmit engine for a different application. An application may request a specific number of each type of engine, depending on the nature of its input/output traffic. The number of receive or transmit engines can be dynamically increased or decreased without suspending or rebooting the host. A universal DMA architecture provides a unified software framework, thereby decreasing the complexity of the software and the hardware gate count cost.
1. A universal DMA (Direct Memory Access) engine architecture for processing either receive or transmit communications, wherein a DMA engine configured according to the architecture comprises:
a software interface for enabling software operating on a host to manage the DMA engine;
a client interface for interacting with either a receive client or a transmit client;
a data plane configured to perform direct memory access for either receive or transmit communications; and
a separate software-posted descriptor ring that is statically and uniquely bound to the DMA engine and, if the DMA engine is configured as a receiver, a hardware-posted descriptor ring bound to the DMA engine, wherein the hardware-posted descriptor ring is shared between the DMA engine and any other DMA engines that are configured as receivers, and wherein the hardware-posted descriptor ring is not bound to the DMA engine if the DMA engine is configured as a transmitter;
a control plane configured to fetch to a memory for the DMA engine software-posted descriptors from the software-posted descriptor ring in a memory for the host and to post hardware-posted descriptors to the hardware-posted descriptor ring;
wherein the software-posted descriptors include transmit packet and receive buffer information and the hardware-posted descriptors include receive packet information.
2. The universal DMA engine architecture of claim 1 , the DMA engine further comprising:
a status plane configured to report the status of the DMA engine to software.
3. The universal DMA engine architecture of claim 2 , wherein said status plane is further configured to detect the status of the receive client or transmit client.
4. The universal DMA engine architecture of claim 1 , wherein:
said universal DMA is dynamically constructed from a pool of DMA resources and allocated to a first application; and
said universal DMA is returned to said pool for reallocation to a second application when the first application terminates.
5. A universal DMA (Direct Memory Access) engine architecture for dynamically constructing a receive or transmit DMA engine from a pool of components by binding components in the pool of components to a DMA engine based on whether the DMA engine is configured as a receive or transmit DMA engine, the pool of components comprising;
a fetch data agent configured to issue data requests to a host computer;
a fetch response agent configured to receive the requested data from the host computer, for transmission via a transmit DMA engine;
a post data agent configured to post to the host computer data received via a receive DMA engine; and
a service arbiter configured to arbitrate between the transmit DMA engine and the receive DMA engine;
wherein the fetch response agent is configured to fetch to a memory for the DMA engine software-posted descriptor from a software-posted descriptor ring in a memory for the host computer, wherein the software-posted descriptor ring is a separate descriptor ring that is statically and uniquely bound to the DMA engine;
wherein the DMA engine is configured to post hardware-posted descriptors to a hardware-posted descriptor ring, wherein the hardware-posted descriptor ring is shared between the DMA engine and any other DMA engines that are configured as receivers, and wherein the hardware-posted descriptor ring is not bound to any DMA engine that is configured as a transmitter; and
wherein the software-posted descriptors include transmit packet and receive buffer information and the hardware-posted descriptors include receive packet information.
6. The universal DMA engine architecture of claim 5 , further comprising:
a post response agent configured to receive acknowledgement of the posted data from the host computer.
7. The universal DMA engine architecture of claim 5 , wherein a DMA engine configured according to the architecture comprises:
a prefetch and cache control module configured to:
issue a data request to the host computer;
receive the requested data using a software descriptor ring; and
cache the requested data for transmission via a transmit client; and
a post and shadow control module configured to:
receive data from a communication link; and
post the received data to the host computer using a receive completion ring.
8. A method of configuring a DMA (Direct Memory Access) engine within a universal DMA architecture, the method comprising:
receiving a request from a first application program operating on a host for at least one DMA engine to perform input/output;
in response to the request, dynamically constructing one or more DMA engines from a pool of universal DMA components configurable as either receive or transmit DMA engines, wherein the universal DMA components include a memory for the DMA engine, wherein constructing the DMA engine includes:
binding a separate software-posted descriptor ring to the DMA engine, wherein the software-posted descriptor ring is statically and uniquely bound to the DMA engine and,
if the DMA engine is configured as a receiver, binding a hardware-posted descriptor ring to the DMA engine, wherein the hardware-posted descriptor ring is shared between the DMA engine and any other DMA engines that are configured as receivers, and wherein the hardware-posted descriptor ring is not bound to the DMA engine if the DMA engine is configured as a transmitter;
allocating the DMA engines to the first application program; and
binding the DMA engines to communication ports.
9. The method of claim 8 , further comprising:
de-allocating the one or more DMA engines;
returning the components of the DMA engines to the pool; and
dynamically constructing one or more other DMA engines from the returned components.
10. A non transitory computer readable medium storing instructions that, when executed by a computer, cause the computer to perform a method of configuring a DMA (Direct Memory Access) engine within a universal DMA architecture, the method comprising:
receiving a request from a first application program for at least one DMA engine to perform input/output;
in response to the request, dynamically constructing one or more DMA engines from a pool of universal DMA components configurable as either receive or transmit DMA engines, wherein the universal DMA components include a memory for the DMA engine, wherein constructing the DMA engine includes:
binding a separate software-posted descriptor ring to the DMA engine, wherein the software-posted descriptor ring is statically and uniquely bound to the DMA engine and,
if the DMA engine is configured as a receiver, binding a hardware-posted descriptor ring to the DMA engine, wherein the hardware-posted descriptor ring is shared between the DMA engine and any other DMA engines that are configured as receivers, and wherein the hardware-posted descriptor ring is not bound to the DMA engine if the DMA engine is configured as a transmitter;
allocating the DMA engines to the first application program; and
binding the DMA engines to communication ports.
11. the universal DMA engine architecture of claim 1 , wherein the DMA engine is dynamically assembled for a software application operating on the host;
wherein the software posts the software-posted descriptors to the descriptor ring; and
wherein the software reads the hardware-posted descriptors from the descriptor ring for processing received packets.
12. The universal DMA engine architecture of claim 11 , wherein the software operating on the host manages a pointer for writing the software-posted descriptors to the descriptor ring, and a pointer for reading the hardware-posted descriptors from the descriptor ring; and
wherein the control plane manages a pointer for reading the software-posted descriptors from the descriptor ring, and a pointer for writing the hardware-posted descriptors to the descriptor ring.
13. The method of claim 8 , further comprising:
receiving a request from a second application program for at least one DMA engine to perform input/output;
in response to the request received from the second application program, de-allocating one or more of the DMA engines for the first application program, and returning components of the de-allocated DMA engines to the pool;
dynamically constructing a DMA engine for the second application program from the returned components, wherein the returned components include resources for a completion ring, wherein the completion ring includes hardware-posted descriptors.