[WIP 0/3] Memory model and atomic API in Rust

[Linus Torvalds]

On Mon, 8 Apr 2024 at 11:14, Al Viro <viro at zeniv.linux.org.uk> wrote:
>
> FWIW, PA-RISC is no better - the same "fetch and replace with constant"
> kind of primitive as for sparc32, only the constant is (u32)0 instead
> of (u8)~0.  And unlike sparc64, 64bit variant didn't get better.

Heh. The thing about PA-RISC is that it is actually *so* much worse
that it was never useful for an arithmetic type.

IOW, the fact that sparc used just a byte meant that the aotmic_t
hackery on sparc still gave us 24 useful bits in a 32-bit atomic_t.

So long ago, we used to have an arithmetic atomic_t that was 32-bit on
all sane architectures, but only had a 24-bit range on sparc.

And I know you know all this, I'm just explaining the horror for the audience.

On PA-RISC you couldn't do that horrendous trick, so parist just used
the "we use a hashed spinlock for all atomics", and "atomic_t" was a
regular full-sized integer type.

Anyway, the sparc 24-bit atomics were actually replaced by the PA-RISC
version back twenty years ago (almost to the day):

   https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git/commit/?id=373f1583c5c5

and while we still had some left-over of that horror in the git tree
up until 2011 (until commit 348738afe530: "sparc32: drop unused
atomic24 support") we probably should have made the
"arch_atomic_xyz()" ops work on generic types rather than "atomic_t"
for a long long time, so that you could use them on other things than
"atomic_t" and friends.

You can see the casting horror here, for example:

   include/asm-generic/bitops/atomic.h

where we do that cast from "volatile unsigned long *p" to
"atomic_long_t *" just to use the raw_atomic_long_xyz() operations.

It would make more sense if the raw atomics took that "native"
volatile unsigned long pointer directly.

(And here that "volatile" is not because it's necessary used as a
volatile - it is - but simply because it's the most permissive type of
pointer. You can see other places using "const volatile unsigned long"
pointers for the same reason: passing in a non-const or non-volatile
pointer is perfectly fine).

              Linus