Re: [PATCH] secretmem: Prevent secretmem_users from wrapping to zero

[Linus Torvalds]

On Mon, Oct 25, 2021 at 2:04 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> Is secretmem different? We're trying to count how many of these we have,
> this is a common pattern in, for example, the network code which does
> this kind of thing a lot.

Yes, secretmem is different.

A refcount being zero means that the data it referenced no longer exists.

That's not what the secretmem counter meant at all.

Making it a refcount was simply WRONG. It's not a refcount, it's a
usage count, and the fact that syzbot caught the warning just shows
how wrong it was.

Stop arguing for garbage. It was wrong, just admit it. The semantics
for "refcount" is something else than what that code had. As a result,
it got reverted. I've applied Willy's patch that actually makes sense.

Arguing for garbage in the name of "security" is still garbage. In
fact, it only causes confusion, and as such is likely to result in
problems - including security problems - later.

Because confusion about semantics is bad. And that was what that patch was.

And I want to state - again - how dangerous this "refcounts are always
prefereable to atomics" mental model is. Refcounts are _not_
fundamentally preferable to atomics. They are slower, bigger, and have
completely different semantics.

So if something isn't a refcount, it damn well shouldn't use "refcount_t".

             Linus