Re: [PATCH RESEND v7 4/6] arm64: futex: refactor futex atomic operation

[Will Deacon <will@kernel.org>]

On Mon, Sep 15, 2025 at 08:40:33PM +0100, Catalin Marinas wrote:
> On Mon, Sep 15, 2025 at 11:32:39AM +0100, Yeoreum Yun wrote:
> > > On Sat, Aug 16, 2025 at 04:19:27PM +0100, Yeoreum Yun wrote:
> > > > diff --git a/arch/arm64/include/asm/futex.h b/arch/arm64/include/asm/futex.h
> > > > index bc06691d2062..ab7003cb4724 100644
> > > > --- a/arch/arm64/include/asm/futex.h
> > > > +++ b/arch/arm64/include/asm/futex.h
> > > > @@ -7,17 +7,21 @@
> > > >
> > > >  #include <linux/futex.h>
> > > >  #include <linux/uaccess.h>
> > > > +#include <linux/stringify.h>
> > > >
> > > >  #include <asm/errno.h>
> > > >
> > > > -#define FUTEX_MAX_LOOPS	128 /* What's the largest number you can think of? */
> > > > +#define LLSC_MAX_LOOPS	128 /* What's the largest number you can think of? */
> > >
> > > I just noticed - you might as well leave the name as is here, especially
> > > if in patch 6 you align down address and use CAS on a 64-bit value as
> > > per Will's comment (and it's no longer LLSC). I think renaming this is
> > > unnecessary.
> > 
> > Okay. I'll restore to use origin name.
> > But I think LSUI wouldn't be used with CAS according to patch 6's
> > comments from you and additionally i think
> > chaning the CAS would make a failure because of
> > change of unrelated field. i.e)
> > 
> > struct user_structure{
> >   uint32 futex;
> >   uint32 some_value;
> > };
> > 
> > In this case, the change of some_value from user side could make a
> > failure of futex atomic operation.
> 
> Yes but the loop would read 'some_value' again, fold in 'futex' and
> retry. It would eventually succeed or fail after 128 iterations if the
> user keeps changing that location. Note that's also the case with LL/SC,
> the exclusive monitor would be cleared by some store in the same cache
> line (well, depending on the hardware implementation) and the STXR fail.
> From this perspective, CAS has better chance of succeeding.
> 
> > So I think it would be better to keep the current LLSC implementation
> > in LSUI.
> 
> I think the code would look simpler with LL/SC but you can give it a try
> and post the code sample here (not in a new series).

If you stick the cas*t instruction in its own helper say, cmpxchg_user(),
then you can do all the shifting/masking in C and I don't reckon it's
that bad. It means we (a) get rid of exclusives, which is the whole
point of this and (b) don't have to mess around with PAN.

> BTW, is there a test suite for all the futex operations? The cover
> letter did not mention any.

I was thinking that too. I'm sure I remember a 'futextest' kicking
around when we did the arm64 port but nowadays there's something in
tools/testing/selftests/futex/ which might be better.

Will