On Mon, Feb 6, 2023, at 13:48, Peter Zijlstra wrote:
> On Mon, Feb 06, 2023 at 01:14:28PM +0100, Peter Zijlstra wrote:
>> On Mon, Feb 06, 2023 at 12:24:00PM +0100, Peter Zijlstra wrote:
>> Basically, using 64bit percpu ops on 32bit is already somewhat dangerous
>> -- wiring up native cmpxchg64 support in that case seemed an
>> improvement.
>>
>> Anyway... let me get on with doing explicit
>> {raw,this}_cpu_cmpxchg{64,128}() thingies.
>
> I only converted x86 and didn't do the automagic downgrade...
>
> Opinions?
I think that's much better, it keeps the interface symmetric
between cmpxchg and this_cpu_cmp_cmpxchg, and makes it harder
to run into the subtle corner case on old x86 CPUs.
For the M486/M586/MK8/MPSC/MATOM/MCORE2 configs, I would
probably want to go with a compile-time check and have most
distro kernels build with at least M586TSC for 32-bit or
an upgraded CONFIG_GENERIC_CPU for x86_64 that assumes cmpxchg16b
(nehalem, second-geneneration k8, silverthorne, ...) in
order to turn system_has_cmpxchg* into a constant value on
all architectures. That can be a separate discussion though
and shouldn't block your series as you just keep the current
state.
> -#ifdef system_has_freelist_aba
> +#ifdef syste_has_freelist_aba
Typo
Arnd
