Direct Memory Access (DMA)
Overview
Direct Memory Access (Controller) is a commonly provided type of co-processor that can typically offload transferring data to and from peripherals and memory.
The DMA API is not a portable API and really cannot be as each DMA has unique memory requirements, peripheral interactions, and features. The API in effect provides a union of all useful DMA functionality drivers have needed in the tree. It can still be a good abstraction, with care, for peripheral devices for vendors where the DMA IP might be very similar but have slight variances.
Driver Implementation Expectations
Synchronization and Ownership
From an API point of view, a DMA channel is a single-owner object, meaning the drivers should not attempt to wrap a channel with kernel synchronization primitives such as mutexes or semaphores. If DMA channels require mutating shared registers, those register updates should be wrapped in a spin lock.
This enables the entire API to be low-cost and callable from any call context, including ISRs where it may be very useful to start/stop/suspend/resume/reload a channel transfer.
Transfer Descriptor Memory Management
Drivers should not attempt to use heap allocations of any kind. If object pools are needed for transfer descriptors then those should be setup in a way that does not break the promise of ISR-allowable calls. Many drivers choose to create a simple static descriptor array per channel with the size of the descriptor array adjustable using Kconfig.
Channel State Machine Expectations
DMA channels should be viewed as state machines that the DMA API provides transition events for in
the form of API calls. Every driver is expected to maintain its own channel state tracking. The busy
state of the channel should be inspectable at any time with dma_get_status().
A diagram showing those expectated possible state transitions and their API calls is provided here for reference.
DMA state finite state machine