A device that moves data by itself can only generate addresses of a certain width: an old card may only reach the first four gigabytes. Before allocating any memory for it, the driver announces that limit, and the kernel then only hands out buffers the device can actually address, using a bounce buffer if it must. Forgetting this in C means the allocator happily returns high memory, the device writes to a truncated address, and something unrelated is corrupted with no clue where it came from. In Rust the call still has to be made; what you gain is an error you must handle if the limit cannot be met.
A device that fetches memory on its own can only form addresses of a limited width, so the driver must announce that width before asking for any buffers. The kernel code shows the calls that do it, m…