Re: [PATCH] doc: Update wake_up() & co. memory-barrier guarantees

[Andrea Parri]

> So I'm not actually sure how many people rely on the RCsc transitive
> smp_mb() here. People certainly rely on the RELEASE semantics, and the
> code itself requires the store->load ordering, together that gives us
> the smp_mb() because that's simply the only barrier we have.
> 
> And looking at smp_mb__after_spinlock() again, we really only need the
> RCsc thing for rq->lock, not for the wakeups. The wakeups really only
> need that RCpc RELEASE + store->load thing (which we don't have).
> 
> So yes, smp_mb(), however the below still makes more sense to me, or am
> I just being obtuse again?

While trying to integrate these remarks into v1 and looking again at the
comment before smp_mb__after_spinlock(), I remembered a discussion where
Boqun suggested some improvements for this comment: so I wrote the commit
reported at the end of this email.

This raises the following two issues:

 1) First, the problem of integrating the resulting comment into v1,
    where I've been talking about _full_ barriers associated to the
    wakeups fuctions but where these are actually implemented as:
    
       spin_lock(s);
       smp_mb__after_spinlock();

 2) Second, the problem of formalizing the requirements described in
    that comment (remark: informally, the LKMM currently states that
    the sequence "spin_lock(s); smp_mb__after_spinlock();" generates
    a full barrier; in particular, this sequence orders
    
       {STORE,LOAD} -> {STORE,LOAD}

    according to the current LKMM).

For (1), I could simply replace each occurrence of "executes a full memory
barrier" with "execute the sequence spin_lock(s); smp_mb__after_spinlock()";
I haven't really thought about (2) yet, but please notice that something as
simple as

   let mb = [...]  |
            ([W] ; po? ; [LKW] ; fencerel(After-spinlock) ; [R])

would _not_ guarantee "RCsc transitivity" ...

A different approach (that could solve both problems at once) would be to
follow the current formalization in LKMM and to modify the comment before
smp_mb__after_spinlock() accordingly (say, informally, "it's required that
that spin_lock(); smp_mb__after_spinlock() provides a full barrier").

Thoughts?

  Andrea