Re: [GIT PULL] x86/shstk for 6.4

[Linus Torvalds]

On Mon, May 15, 2023 at 3:40 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
>         mutex_lock(&mm->fork_lock);
>         if (atomic_read(&mm->mm_users) > 1 ||
>             atomic_read(&mm->mm_count) > 1) {

Again, just to clarify: at this point the mm_users/count checks are
purely a heuristic. If they are both 1, then we know we're *currently*
the only user, and the fork_lock means that no new users that matter
can come in.

If we race with things like /proc accesses that have raised either of
those counts, and we happen to say "ok, we're not alone in the
universe", then that just means that we fall back to the slow
thread-safe path. Unfortunate, but not a huge deal.

The slow path is what we do right now, and while it might be a bit
slower with a 'lock cmpxchg' instead of a 'lock andb', it shouldn't be
some *insanely* much slower case. So what we really want to do here is
an optimistic and cheap "can we do the fast case" for the presumably
common situation where we're really just a bog-standard
single-threaded fork.

IOW, the mm counts aren't somehow semantically important, we're
literally just looking for a cheap test for a common case.

That's my argument, at least. There may be flaws in that argument,
like some memory ordering with some exit() case that decrements
mm_users, but those should very much happen only after the exit() has
already stopped using the page tables, so I don't think those can
actually matter.

Keyword here being "I don't think".

               Linus