On Thu, 26 Oct 2023 at 08:36, Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote: > > > asm volatile("xchg %b0,%1" > > which has an implicit lock prefix (xchg with a memory operand is a > special-case): Yeah, this is why we do "percpu_xchg()" - which does not want locked semantics - as a "cmpxchg" loop. Steven, check out arch/x86/include/asm/percpu.h for a rough implementation of a 'xchg()' without SMP coherency, just cpu-local one (ie atomic wrt being preempted by the kernel, but not atomic wrt other CPU's accessing the same variable concurrently) Linus