I am developing a driver for a DMA bus master device, part of an SoC powered by a Cortex M7 CPU. Suppose I have two memory locations, x and y, which map to the same cache line, which is normal, write-back cacheable memory, and suppose the following sequence of events:1. Start with x = x1, y = y1, cache line invalid.2. CPU reads y3. DMA device sets x = x2, in memory4. CPU sets y = y25. CPU cleans the cache line.After 5. completes, from the point of view of the DMA device, x = ?I think the DMA will see x = x1, here is my reasoning:- When CPU reads y in 2., the cache line gets pulled in cache. It reads x = x1, y = y1, and is marked as valid.- The DMA then updates x in memory, but the change is not reflected in the cache line.- When the CPU sets y = y2, the cache line is marked as dirty.- When the CPU cleans the cache line, as it is dirty it gets written back to memory.- When it gets written back to memory, it reads x = x1, y = y2, thus overwriting thechange made by the DMA to x.Does that sound like a good reasoning?
x
y
x = x1, y = y1
x = x2
y = y2
x = ?
x = x1
x = x1, y = y2
Correct. The "reading for y" kinda neglects the invalidation. So you should at best have no other data in a memory area which is written by the DMA.
Thank you for your reply. So assuming I keep descriptors in cacheable memory, if I want to invalidate / clean descriptors individually, I need to make sure each of them lies in a different cache line, and also make sure I don't overwrite any other data structure in the same cache line, also written by the DMA?
descriptors are (most often) not the size of a cache line, so it is better to place them in non-cache ("normal") memory. The pain of invalidating/cleaning cacheline because of DMA change it is not worth the effort.