On Fri, Jul 23, 2021 at 06:07:44PM +0200, Arnd Bergmann wrote:

There is also a problem with locking: In the case that both firmware and
kernel have to access registers within a syscon area, you may need to
have a semaphore to protect an atomic sequence of accesses, but since
the interface only provides a single register load/store, there is no way for
a kernel driver to serialize against a firmware-internal driver.

The standard solution to this for the read/modify/write case would be to
expose an explicit update_bits() operation (some hardware does this for
concurrency and/or bus bandwidth/latency reasons), though that doesn't
help with larger or multi-register sequences (and to be clear as I've
been saying I don't think we should do this at all).