Re: [PATCH v2 1/2] locking: Implement an algorithm choice for Wound-Wait mutexes

[Thomas Hellstrom]

On 06/14/2018 02:48 PM, Thomas Hellstrom wrote:
> Hi, Peter,
>
> On 06/14/2018 02:41 PM, Peter Zijlstra wrote:
> > On Thu, Jun 14, 2018 at 09:29:21AM +0200, Thomas Hellstrom wrote:
> > > +static bool __ww_mutex_wound(struct mutex *lock,
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ struct ww_acquire_ctx *ww_ctx,
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ struct ww_acquire_ctx *hold_ctx)
> > > +{
> > > +ÂÂÂ struct task_struct *owner = __mutex_owner(lock);
> > > +
> > > +ÂÂÂ lockdep_assert_held(&lock->wait_lock);
> > > +
> > > +ÂÂÂ if (owner && hold_ctx && __ww_ctx_stamp_after(hold_ctx, ww_ctx) &&
> > > +ÂÂÂÂÂÂÂ ww_ctx->acquired > 0) {
> > > +ÂÂÂÂÂÂÂ hold_ctx->wounded = 1;
> > > +
> > > +ÂÂÂÂÂÂÂ /*
> > > +ÂÂÂÂÂÂÂÂ * wake_up_process() paired with set_current_state() inserts
> > > +ÂÂÂÂÂÂÂÂ * sufficient barriers to make sure @owner either sees it's
> > > +ÂÂÂÂÂÂÂÂ * wounded or has a wakeup pending to re-read the wounded
> > > +ÂÂÂÂÂÂÂÂ * state.
> > > +ÂÂÂÂÂÂÂÂ *
> > > +ÂÂÂÂÂÂÂÂ * The value of hold_ctx->wounded in
> > > +ÂÂÂÂÂÂÂÂ * __ww_mutex_lock_check_stamp();
> > > +ÂÂÂÂÂÂÂÂ */
> > > +ÂÂÂÂÂÂÂ if (owner != current)
> > > +ÂÂÂÂÂÂÂÂÂÂÂ wake_up_process(owner);
> > > +
> > > +ÂÂÂÂÂÂÂ return true;
> > > +ÂÂÂ }
> > > +
> > > +ÂÂÂ return false;
> > > +}
> > > @@ -338,12 +377,18 @@ ww_mutex_set_context_fastpath(struct ww_mutex 
> > > *lock, struct ww_acquire_ctx *ctx)
> > > ÂÂÂÂÂÂ * and keep spinning, or it will acquire wait_lock, add itself
> > > ÂÂÂÂÂÂ * to waiter list and sleep.
> > > ÂÂÂÂÂÂ */
> > > -ÂÂÂ smp_mb(); /* ^^^ */
> > > +ÂÂÂ smp_mb(); /* See comments above and below. */
> > > Â ÂÂÂÂÂ /*
> > > -ÂÂÂÂ * Check if lock is contended, if not there is nobody to wake up
> > > +ÂÂÂÂ * Check if lock is contended, if not there is nobody to wake up.
> > > +ÂÂÂÂ * We can use list_empty() unlocked here since it only compares a
> > > +ÂÂÂÂ * list_head field pointer to the address of the list head
> > > +ÂÂÂÂ * itself, similarly to how list_empty() can be considered 
> > > RCU-safe.
> > > +ÂÂÂÂ * The memory barrier above pairs with the memory barrier in
> > > +ÂÂÂÂ * __ww_mutex_add_waiter and makes sure lock->ctx is visible 
> > > before
> > > +ÂÂÂÂ * we check for waiters.
> > > ÂÂÂÂÂÂ */
> > > -ÂÂÂ if (likely(!(atomic_long_read(&lock->base.owner) & 
> > > MUTEX_FLAG_WAITERS)))
> > > +ÂÂÂ if (likely(list_empty(&lock->base.wait_list)))
> > > ÂÂÂÂÂÂÂÂÂ return;
> > OK, so what happens is that if we see !empty list, we take wait_lock,
> > if we end up in __ww_mutex_wound() we must really have !empty wait-list.
> >
> > It can however still see !owner because __mutex_unlock_slowpath() can
> > clear the owner field. But if owner is set, it must stay valid because
> > FLAG_WAITERS and we're holding wait_lock.
>
> If __ww_mutex_wound() is called from ww_mutex_set_context_fastpath() 
> owner is the current process so we can never see !owner. However if 
> __ww_mutex_wound() is called from __ww_mutex_add_waiter() then the 
> above is true.

Or actually it was intended to be true, but FLAG_WAITERS is set too 
late. It needs to be moved to just after we actually add the waiter to 
the list.

Then the hunk that replaces a FLAG_WAITERS read with a lockless 
list_empty() can also be ditched.

/Thomas


> > So the wake_up_process() is in fact safe.
> >
> > Let me put that in a comment.
>
>
> Thanks,
>
> Thomas