Investigate options for exposing MCTP to userspace
amboar opened this issue · comments
We need to develop a sound story for supporting MCTP in the kernel. We need to determine how much we should do in userspace vs the kernel. With libmctp it's possible that we could do a lot of the work in userspace with the development effort amortised across the community of libmctp consumers, but my concern with this approach is the number of disparate kernel interfaces we might need to consume.
Investigate whether moving the handling of MCTP bindings into the kernel will help unify and minimise the amount of effort required in userspace. The motivating concern here is around OpenPOWER's proposed LPC MCTP binding which makes use of multiple elements on the LPC bus (a KCS device in LPC IO space, and a portion of the LPC FW space).
MCTP describes what is, in effect, a very simple network stack. Investigate whether using an AF_MCTP socket is a useful abstraction. This may depend on how we want to handle bridging of multiple network segments, and whether we want to perform packet routing in the kernel or userspace.
We've decided that socket-based MCTP is where we want to go.