On Fri, Dec 02, 2022 at 03:01:58PM +0000, Mel Gorman wrote:
> On Fri, Dec 02, 2022 at 12:21:06PM +0100, Sebastian Andrzej Siewior wrote:
> > On 2022-12-02 10:02:23 [+0000], Mel Gorman wrote:
> > > The lock owner is updated with an IRQ-safe raw spinlock held but the
> > > spin_unlock does not provide acquire semantics which are needed when
> > > acquiring a mutex. This patch adds the necessary acquire semantics for a
> > > lock operation when the lock owner is updated. It successfully completed
> > > 10 iterations of the dbench workload while the vanilla kernel fails on
> > > the first iteration.
> >
> > I *think* it is
> >
> > Fixes: 700318d1d7b38 ("locking/rtmutex: Use acquire/release semantics")
> >
>
> Adding Davidlohr to cc.
>
> It might have made the problem worse but even then rt_mutex_set_owner was
> just a plain assignment and while I didn't check carefully, at a glance
> try_to_take_rt_mutex didn't look like it guaranteed ACQUIRE semantics.
>
> > Before that, it did cmpxchg() which should be fine.
> >
> > Regarding mark_rt_mutex_waiters(). Isn't acquire semantic required in
> > order for the lock-owner not perform the fastpath but go to the slowpath
> > instead?
> >
>
> Good spot, it does. While the most straight-forward solution is to use
> cmpxchg_acquire, I think it is overkill because it could incur back-to-back
> ACQUIRE operations in the event of contention. There could be a smp_wmb
> after the cmpxchg_relaxed but that impacts all arches and a non-paired
> smp_wmb is generally frowned upon.
>
> I'm thinking this on top of the patch should be sufficient even though
> it's a heavier operation than is necesary for ACQUIRE as well as being
> "not typical" according to Documentation/atomic_t.txt. Will, as this
> affects ARM primarily do you have any preference?
>
> diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c
> index 35212f260148..af0dbe4d5e97 100644
> --- a/kernel/locking/rtmutex.c
> +++ b/kernel/locking/rtmutex.c
> @@ -238,6 +238,13 @@ static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
> owner = *p;
> } while (cmpxchg_relaxed(p, owner,
> owner | RT_MUTEX_HAS_WAITERS) != owner);
> +
> + /*
> + * The cmpxchg loop above is relaxed to avoid back-to-back ACQUIRE
> + * operations in the event of contention. Ensure the successful
> + * cmpxchg is visible.
> + */
> + smp_mb__after_atomic();
Could we use smp_acquire__after_ctrl_dep() instead?
Will
