On Fri, Oct 16, 2020 at 12:53:13PM +0200, Peter Zijlstra wrote: > That's like saying: "I'm too lazy to track what I've looked at already". > You're basically proposing to graffiti "Kees was here -- 16/10/2020" all > over the kernel. Just so you can see where you still need to go. > > It says the code was (assuming your audit was correct) good at that > date, but has no guarantees for any moment after that. That kind of bit-rot marking is exactly what I would like to avoid: just putting a comment in is pointless. Making the expectations of the usage become _enforced_ is the goal. And having it enforced by the _compiler_ is key. Just adding a meaningless attribute that a static checker will notice some time and hope people fix them doesn't scale either (just look at how many sparse warnings there are). So with C's limits, the API and type enforcement become the principle way to get this done. So, if there are behavioral patterns we CAN carve away from atomic_t cleanly (and I think there are), those are the ones I want to work towards. The "corner case" uses of atomic_t are much less common than the "big" code patterns like lifetime management (now delegated to and enforced by refcount_t). My estimation was that _statistics_ (and not "serial identifiers") was the next biggest code pattern using atomic_t that could benefit from having its usage isolated. It is not itself a dangerous code pattern, but it can mask finding them. Then, at the end of the day, only the corner cases remain, and those can be seen clearly as they change over time. Since we can never have a one-time audit be anything other than advisory, we need to make it EASY to do those kinds of audits so they can be done regularly. -- Kees Cook
