Re: [PATCH 2/3] sparc32: emulate casa on V8 CPUs

[Imre Kaloz <kaloz@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 2142 bytes --]

Hi Ramin,

On Thu, 15 Oct 2026, Ramin Moussavi wrote:

> On Sun, Sep 27, 2026 at 02:58 AM Imre Kaloz wrote:
>> so a C library can use one casa for its 32-bit atomics on any V8 part.
>
> I sent a series covering part of the same ground, Magnus and Adrian
> pointed me here -- I had not seen yours, so we worked on the same
> problem in parallel:
> https://lore.kernel.org/sparclinux/20261015085545.3246951-1-ramin.moussavi@yacoub.de/
>
> Your casa emulation looks like a good approach to me. What I would
> still need is a way for a C library to find out whether the kernel
> has it: a library built for v8 needs to keep its own fallback for
> older kernels, and there casa is SIGILL. I see three ways a library
> could tell:
>
> - an AT_HWCAP bit, like HWCAP_SPARC_CASTRAP in Magnus' RFC but meaning
>  "casa works, in hardware or emulated",
> - a system call it can probe for ENOSYS, like patch 4 of my series,
> - or trying casa once under a SIGILL handler at startup.
>
> I know glibc does not want arch-specific atomics: on v8 the compiler
> turns the atomic builtins into calls, and glibc would rather have
> libgcc implement those with casa. But libgcc then faces the same
> question, unless it simply requires a kernel with the emulation; and
> uClibc-ng in my case keeps a fallback for older kernels. So from the
> kernel side the question how a library can tell remains.

Thanks, glad the approach works for you too.

My own view is that moving a C library to casa means rebuilding it
anyway, and that build can simply require a kernel with the
emulation, as with any other new kernel feature a library starts to
depend on. On an older kernel it stops with SIGILL at the first
casa, which at least fails clearly. A library that wants to keep its
fallback can probe casa once under a SIGILL handler, and that works
on every kernel without new ABI. Whichever way you go, use ASI 0x0A:
it also runs natively on LEON parts that have casa.

Whether the kernel should offer an explicit way to tell is Andreas’
call, though. Andreas, if you’d like one, I’m happy to add whatever
you prefer, an AT_HWCAP bit or something else.


Thanks,
Imre