Re: [RFC PATCH] riscv/locking: Strengthen spin_lock() and spin_unlock()

[Will Deacon <will.deacon@arm.com>]

On Mon, Feb 26, 2018 at 08:06:59AM -0800, Linus Torvalds wrote:
> On Mon, Feb 26, 2018 at 6:21 AM, Luc Maranget <luc.maranget@inria.fr> wrote:
> >
> > That is, locks are not implemented from more basic primitive but are specified.
> > The specification can be described as behaving that way:
> >   - A lock behaves as a read-modify-write. the read behaving as a read-acquire
> 
> This is wrong, or perhaps just misleading.
> 
> The *whole* r-m-w acts as an acquire. Not just the read part. The
> write is very much part of it.
> 
> Maybe that's what you meant, but it read to me as "just the read part
> of the rmw behaves as a read-acquire".
> 
> Because it is very important that the _write_ part of the rmw is also
> ordered wrt everything that is inside the spinlock.
> 
> So doing a spinlock as
> 
>  (a) read-locked-acquire
>    modify
>  (c) write-conditional
> 
> would be wrong, because the accesses inside the spinlock are ordered
> not just wrt the read-acquire, they have to be ordered wrt the write
> too.
> 
> So it is closer to say that it's the _write_ of the r-m-w sequence
> that has the acquire semantics, not the read.

Strictly speaking, that's not what we've got implemented on arm64: only
the read part of the RmW has Acquire semantics, but there is a total
order on the lock/unlock operations for the lock. For example, if one
CPU does:

spin_lock(&lock);
WRITE_ONCE(foo, 42);

then another CPU could do:

if (smp_load_acquire(&foo) == 42)
	BUG_ON(!spin_is_locked(&lock));

and that could fire. Is that relied on somewhere?

Will