How should Linux and a BeagleBone PRU exchange blocks without overrunning each other?

Give each block explicit ownership and a sequence number. The producer should publish completion only after the block is ready, and the consumer should release it only after finishing. Apply the platform's memory-ordering and visibility requirements rather than relying on a shared flag alone.

Size the buffer for a measured interruption budget. At 100,000 samples per second and two bytes per sample, 100 ms of backlog requires 20,000 bytes before metadata. Extra capacity handles a bounded delay; it cannot fix a consumer that is slower on average indefinitely.

Define an overrun policy and expose counters to the application. Test storage stalls, CPU load and host-process restarts. If the data must be lossless, prove that the sustained path and worst expected stalls fit the budget, or provide an explicit backpressure mechanism the source can obey.