Re: are volatile and memory barriers necessary for single threaded code?

[Lorenzo Gabriele <lorenzolespaul@gmail.com>]

Thank you very much for the detailed answer!
So I think I will more safely build on top of liburing!

Lorenzo

Il giorno mer 6 mag 2020 alle ore 15:45 Pavel Begunkov
<asml.silence@gmail.com> ha scritto:
>
> On 04/05/2020 19:54, Lorenzo Gabriele wrote:
> > Hi everyone,
> > I'm a complete noob so sorry if I'm saying something stupid.
> > I want to have a liburing-like library for the Scala Native language.
> > I can't easily use liburing itself because of some limitations of the
> > language.. So I was rewriting the C code in liburing in Scala Native.
> > The language is single threaded and, sadly, doesn't support atomic,
> > nor volatile. I was thinking what are the implications of completely
> > removing the memory barriers.
> > Are they needed for something related with multithreading or they are
> > needed regardless to utilize io_uring?
>
> Long story short, even if your app is single-threaded, io_uring is _not_.
> I wouldn't recommend removing it. See the comment below picked from io_uring.h
>
> /*
>  * After the application reads the CQ ring tail, it must use an
>  * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
>  * before writing the tail (using smp_load_acquire to read the tail will
>  * do). It also needs a smp_mb() before updating CQ head (ordering the
>  * entry load(s) with the head store), pairing with an implicit barrier
>  * through a control-dependency in io_get_cqring (smp_store_release to
>  * store head will do). Failure to do so could lead to reading invalid
>  * CQ entries.
>  */
>
>
> More difficult to say, what will actually happen. E.g. if you don't use polling
> io_uring modes, and if you don't do speculative CQ reaping, there is a pairing
> smp_rmb() just before returning from a wait. But, again, the io_uring ABI
> doesn't guarantee correctness without them.
>
> --
> Pavel Begunkov