Re: [Intel-gfx] [PATCH 1/5] linux/minmax.h: add non-atomic version of xchg

From: Daniel Vetter
Date: Thu Jan 05 2023 - 09:13:50 EST


On Thu, Jan 05, 2023 at 01:34:33PM +0000, David Laight wrote:
> From: Jani Nikula
> > Sent: 05 January 2023 13:28
> >
> > On Thu, 05 Jan 2023, Daniel Vetter <daniel@xxxxxxxx> wrote:
> > > On Mon, Dec 12, 2022 at 09:38:12AM +0000, David Laight wrote:
> > >> From: Andrzej Hajda <andrzej.hajda@xxxxxxxxx>
> > >> > Sent: 09 December 2022 15:49
> > >> >
> > >> > The pattern of setting variable with new value and returning old
> > >> > one is very common in kernel. Usually atomicity of the operation
> > >> > is not required, so xchg seems to be suboptimal and confusing in
> > >> > such cases. Since name xchg is already in use and __xchg is used
> > >> > in architecture code, proposition is to name the macro exchange.
> > >>
> > >> Dunno, if it is non-atomic then two separate assignment statements
> > >> is decidedly more obvious and needs less brain cells to process.
> > >> Otherwise someone will assume 'something clever' is going on
> > >> and the operation is atomic.
> > >
> > > Yes, this also my take. The i915 code that uses this to excess is decidely
> > > unreadable imo, and the macro should simply be replaced by open-coded
> > > versions.
> > >
> > > Not moved into shared headers where even more people can play funny games
> > > with it.
> >
> > My stand in i915 has been that the local fetch_and_zero() needs to
> > go. Either replaced by a common helper in core kernel headers, or open
> > coded, I personally don't care, but the local version can't stay.
> >
> > My rationale has been that fetch_and_zero() looks atomic and looks like
> > it comes from shared headers, but it's neither. It's deceptive. It
> > started small and harmless, but things like this just proliferate and
> > get copy-pasted all over the place.

Yeah the entire "is it atomic or not" is the issue on top here.

> > So here we are, with Andrzej looking to add the common helper. And the
> > same concerns crop up. What should it be called to make it clear that
> > it's not atomic? Is that possible?
>
> old_value = read_write(variable, new_value);
>
> But two statements are much clearer.

Yeah this is my point for fetch_and_zero or any of the other proposals.
We're essentially replacing these two lines:

var = some->pointer->chase;
some->pointer->chase = NULL;

with a macro. C is verbose, and sometimes painfully so, if the pointer
chase is really to onerous then I think that should be refactored with a
meaningfully locally name variable, not fancy macros wrapped around to
golf a few characters away.

But what about swap() you ask? That one needs a temp variable, and it does
make sense to hide that in a ({}) block in a macro. But for the above two
lines I really don't see a point outside of obfuscated C contexts.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch