Re: [PATCH 1/1] futex: futex_wake_op, fix sign_extend32 sign bits

From: Darren Hart
Date: Mon Dec 11 2017 - 17:56:18 EST


On Mon, Dec 11, 2017 at 08:37:11AM +0100, Jiri Slaby wrote:
> On 12/10/2017, 09:50 PM, Linus Torvalds wrote:
> > On Thu, Nov 30, 2017 at 6:35 AM, Jiri Slaby <jslaby@xxxxxxx> wrote:
> >> sign_extend32 counts the sign bit parameter from 0, not from 1. So we
> >> have to use "11" for 12th bit, not "12".
> >
> > This interface is crap. It really doesn't make much sense. I wonder
> > how many people have gotten this wrong, but it's hard to tell.
>
> I tend to agree, because it really surprised me. So at that time I
> searched for most (all?) uses of the interface, checked them and all of
> them *seem* to be fine.
>
> > I'm applying this directly to my tree since I didn't see anybody else
> > react to it, but the whole pattern worries me.
> >
> > Also, clearly nobody actually uses the odder corners of futex ops
> > anyway. Maybe we should deprecate them entirely?
> >
> > Jiri, did you notice by testing, or what?
>
> I noticed it by coincidence while fixing the strace build test failures
> -- e78c38f6bdd9 (futex: futex_wake_op, do not fail on invalid op). I
> compiled a bit modified futex_atomic_op_inuser in userspace to test the
> conversion and the added check and it did not work.
>
> And yes, somebody (tglx?) noted already that this interface is old and
> perhaps unused.

The only use I know of for FUTEX_WAKE_OP is glibc
lll_futex_wake_unlock(). and that is limited to a single operation.

At the very least, we need to add a futex_wake_op test to the
kselftests, something that's been nagging me for a very long time.
There are some 120 different combinations of op and cmp and condition
value.

Assuming this isn't urgent, I've added it to my projects list.

--
Darren Hart
VMware Open Source Technology Center