[openib-general] Re: userspace doorbells

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: userspace doorbells
> 
>     Michael> Roland, I see you have made the doorbell page volatile.
>     Michael> This makes sence, and must be enough on x86_64, but for
>     Michael> this to work on PPC, wont you still need to insert a
>     Michael> write memory barrier, to guard against the CPU
>     Michael> re-ordering writes to hardware and to the WQE?  Since you
>     Michael> do it in kernel, why not in userspace?
> 
> I'm working on it... see the <infiniband/arch.h> file I added to
> libibverbs for the start of my plan.
> 
>  - R.
> 

OK, makes sence. I expect you'll need rmb for CQE polling and wmb for
doorbells, just like we do for kernel code.

Hmm, and I expect pthread_spin_lock may provide a generic barrier
implementation: it seems pthread_spin_lock just has to include
an rmb, pthread_spin_unlock - a wmb.

-- 
MST - Michael S. Tsirkin