Re: [RFC][PATCH 2/3] locking: Annotate spin_unlock_wait() users

[Linus Torvalds]

On Tue, May 24, 2016 at 7:27 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> spin_unlock_wait() has an unintuitive 'feature' in that it doesn't
> fully serialize against the spin_unlock() we've waited on.

NAK.

We don't start adding more of this "after_ctrl_dep" crap.

It's completely impossible to understand, and even people who have
been locking experts have gotten it wrong.

So it is *completely* unacceptable to have it in drivers.

This needs to be either hidden inside the basic spinlock functions,
_or_ it needs to be a clear and unambiguous interface. Anything that
starts talking about control dependencies is not it.

Note that this really is about naming and use, not about
implementation. So something like "spin_sync_after_unlock_wait()" is
acceptable, even if the actual _implementation_ were to be exactly the
same as the "after_ctrl_dep()" crap.

The difference is that one talks about incomprehensible implementation
details that nobody outside of the person who *implemented* the
spinlock code is supposed to understand (and seriously, I have my
doubts even the spinlock implementer understands it, judging by the
last time this happened), and the other is a much simpler semantic
guarantee.

So don't talk about "acquire". And most certainly don't talk about
"control dependencies". Not if we end up having things like *drivers*
using this like in this example libata.

                Linus