On Fri, Mar 03, 2017 at 03:49:38PM +0100, Peter Zijlstra wrote:
On Fri, Mar 03, 2017 at 09:26:50AM +0100, Christian Borntraeger wrote:
Right. The main purpose is to read/write _ONCE_. You can assume a somewhat atomic access for sizes <= word size. And there are certainly places that rely on that. But the *ONCE thing is mostly used for things where we used barrier() 10 years ago.
A lot of code relies on READ/WRITE_ONCE() to generate single instructions for naturally aligned machined word sized loads/stores (something GCC used to guarantee, but does no longer IIRC).
So much so that I would say its a bug if READ/WRITE_ONCE() doesn't generate a single instruction under those conditions.
However, every time I've tried to introduce stricter semantics/primitives to verify things Linus hated it.
See here for the last attempt:
https://marc.info/?l=linux-virtualization&m=148007765918101&w=2