Re: [PATCH] rcutorture: Fix rcu_torture_pipe_update_one()/rcu_torture_writer() data race and concurrency bug

From: Linus Torvalds
Date: Wed Mar 06 2024 - 14:02:25 EST


On Wed, 6 Mar 2024 at 10:53, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> Now, are you OK with an addition of ADD_ONCE() and/or INC_ONCE()? So that we
> don't have to look at:
>
> WRITE_ONCE(a, READ_ONCE(a) + 1);
>
> ?

In a generic header file under include/linux/?

Absolutely not. The above is a completely broken operation. There is
no way in hell we should expose it as a general helper.

So there is no way we'd add that kind of sh*t-for-brains operation in
(for example) our <asm/rwonce.h> header file next to the normal
READ/WRITE_ONCE defines.

In some individual tracing C file where it has a comment above it how
it's braindamaged and unsafe and talking about why it's ok in that
particular context? Go wild.

But honestly, I do not see when a ADD_ONCE() would ever be a valid
thing to do, and *if* it's a valid thing to do, why you'd do it with
READ_ONCE and WRITE_ONCE.

If you don't care about races, just do a simple "++" and be done with
it. The end result is random.

Adding a "ADD_ONCE()" macro doesn't make it one whit less random. It
just makes a broken concept even uglier.

So honestly, I think the ADD_ONCE macro not only needs to be in some
tracing-specific C file, the comment needs to be pretty damn big too.
Because as a random number generator, it's not even a very good one.
So you need to explain *why* you want a particularly bad random number
generator in the first place.

Linus