Re: [PATCH v4 5/5] rqspinlock: use smp_cond_load_acquire_timewait()

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Mon, Sep 1, 2025 at 4:28 AM Catalin Marinas <catalin.marinas@arm.com> wrote:
>
> On Fri, Aug 29, 2025 at 01:07:35AM -0700, Ankur Arora wrote:
> > diff --git a/arch/arm64/include/asm/rqspinlock.h b/arch/arm64/include/asm/rqspinlock.h
> > index a385603436e9..ce8feadeb9a9 100644
> > --- a/arch/arm64/include/asm/rqspinlock.h
> > +++ b/arch/arm64/include/asm/rqspinlock.h
> > @@ -3,6 +3,9 @@
> >  #define _ASM_RQSPINLOCK_H
> >
> >  #include <asm/barrier.h>
> > +
> > +#define res_smp_cond_load_acquire_waiting() arch_timer_evtstrm_available()
>
> More on this below, I don't think we should define it.
>
> > diff --git a/kernel/bpf/rqspinlock.c b/kernel/bpf/rqspinlock.c
> > index 5ab354d55d82..8de1395422e8 100644
> > --- a/kernel/bpf/rqspinlock.c
> > +++ b/kernel/bpf/rqspinlock.c
> > @@ -82,6 +82,7 @@ struct rqspinlock_timeout {
> >       u64 duration;
> >       u64 cur;
> >       u16 spin;
> > +     u8  wait;
> >  };
> >
> >  #define RES_TIMEOUT_VAL      2
> > @@ -241,26 +242,20 @@ static noinline int check_timeout(rqspinlock_t *lock, u32 mask,
> >  }
> >
> >  /*
> > - * Do not amortize with spins when res_smp_cond_load_acquire is defined,
> > - * as the macro does internal amortization for us.
> > + * Only amortize with spins when we don't have a waiting implementation.
> >   */
> > -#ifndef res_smp_cond_load_acquire
> >  #define RES_CHECK_TIMEOUT(ts, ret, mask)                              \
> >       ({                                                            \
> > -             if (!(ts).spin++)                                     \
> > +             if ((ts).wait || !(ts).spin++)                \
> >                       (ret) = check_timeout((lock), (mask), &(ts)); \
> >               (ret);                                                \
> >       })
> > -#else
> > -#define RES_CHECK_TIMEOUT(ts, ret, mask)                           \
> > -     ({ (ret) = check_timeout((lock), (mask), &(ts)); })
> > -#endif
>
> IIUC, RES_CHECK_TIMEOUT in the current res_smp_cond_load_acquire() usage
> doesn't amortise the spins, as the comment suggests, but rather the
> calls to check_timeout(). This is fine, it matches the behaviour of
> smp_cond_load_relaxed_timewait() you introduced in the first patch. The
> only difference is the number of spins - 200 (matching poll_idle) vs 64K
> above. Does 200 work for the above?
>
> >  /*
> >   * Initialize the 'spin' member.
> >   * Set spin member to 0 to trigger AA/ABBA checks immediately.
> >   */
> > -#define RES_INIT_TIMEOUT(ts) ({ (ts).spin = 0; })
> > +#define RES_INIT_TIMEOUT(ts) ({ (ts).spin = 0; (ts).wait = res_smp_cond_load_acquire_waiting(); })
>
> First of all, I don't really like the smp_cond_load_acquire_waiting(),
> that's an implementation detail of smp_cond_load_*_timewait() that
> shouldn't leak outside. But more importantly, RES_CHECK_TIMEOUT() is
> also used outside the smp_cond_load_acquire_timewait() condition. The
> (ts).wait check only makes sense when used together with the WFE
> waiting.

+1 to the above.

Penalizing all other architectures with pointless runtime check:

> -             if (!(ts).spin++)                                     \
> +             if ((ts).wait || !(ts).spin++)                \

is not acceptable.