On Thu, 21 Nov 2024 at 19:57, Linus Torvalds
[off-list ref] wrote:

Anyway, that was a long way of saying: I really think we should just
special-case the (few) important cases that get reported. Because any
*big* improvements will come not from just inlining.

Looking around at the futex code some more, I note:

 - the cmpxchg case and futex ops use an explicit barrier too, which is bad

 - we'd actually be better off inlining not just the user access, but
the whole futex_get_value_locked(), because then the compiler will be
able to do CSE on the user address masking, and only do it once
(several places do multiple different futex_get_value_locked() calls).

iow, I think the fix for the futex case ends up being a patch
something like the attached.

[ Annoyingly, we need "can_do_masked_user_access()" even on x86,
because the 32-bit case doesn't do the address masking trick ]

I've only compiled it so far, about to actually boot into it. Pray for me,

               Linus