Bufferalways onlytome

[troelsjessen]

Most commits will be squashed prior to merge

A few commits are not directly related to this improvements, those will go as is.

[troelsjessen]

Johan states:
Jeg synes vi skal lave funktionen i routeren, således at når pakker kommer ind, og buffree er under et vist tal, eksempelvis 5 eller 10, så skal routeren stoppe med at route, og kun acceptere pakker "tome". Resten bliver discarded af routeren. - Det har den fordel at det automatisk virker for alle interfaces, og vi slipper for at lave "lag 3 inspektion" på lag 2.

[troelsjessen]

I just tried Johans approach (which is significantly more elegant), but unfortunately it fails as most TX functions are synchronous and blocking the router thread.
This means that several packets can come in through high-speed interfaces (ZMQ, ETH, SpW etc.), typically having designated RX threads or ISR routines, exhausting the buffer queue while the router is waiting to finish transmission of a single CAN packet.
In the current implementation, buffer_get_always is actually not used for ZMQ and ETH (which is a mistake), but packets can still come through CAN and exhaust the buffer queue.

[johandc]

Aha! That in itself is an issue, so maybe we should quite early separate the routing queue and the incoming queue, and process the tx of routed traffic separately from incoming traffic. Incoming traffic probably doesnt even need a task to process it, because it can usually be placed into the correct queue immediately.

[troelsjessen]

I think Johans latest comment is reaching beyond the current requirements. As Johan states, most interface types rely on a task for for handling incoming messages, if not handled in an ISR.
Nevertheless, TX is currently carried out simply by having the router calling a blocking transmit function going into the interface and further into the driver layer. Demanding these to be asynchronous procedures seem like a significant change to me, likely requiring additional resources from modules hosting a CSP stack.

[johandc]

So what i am suggesting is to drop them before they arrive at the TX queue, in the csp_qfifo_write call.