On Mon, Dec 19, 2022, at 16:35, Peter Zijlstra wrote:
> In order to replace cmpxchg_double() with the newly minted
> cmpxchg128() family of functions, wire it up in this_cpu_cmpxchg().
>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
Does this work on x86 chips without X86_FEATURE_CX16?
As far as I can tell, the new percpu_cmpxchg128_op uses
the cmpxchg16b instruction unconditionally without checking
for the feature bit first, and is now used by this_cpu_cmpxchg()
unconditionally as well.
This is fine for the moment if the only user is mm/slub.c
and that retains the system_has_cmpxchg128() runtime check,
but I think a better interface would be to guarantee that
this_cpu_cmpxchg() always ends up either in a working
inline asm or the generic fallback but not an invalid
opcode.
For consistency, I would also suggest this_cpu_cmpxchg() to
take the same argument types as cmpxchg(): at most 'unsigned
long' sized, with additional this_cpu_cmpxchg64() and
this_cpu_cmpxchg128() macros that take fixed-size arguments.
I have an older patch set that additionally converts all
8-bit and 16-bit cmpxchg()/xchg() calls to cmpxchg_8()/
xchg_8()/cmpxchg_16()/xchg_16() and and leaves only the
fixed-32bit and variable typed 'unsigned long' sized
callers for the weakly typed variant.
Arnd
