Re: [PATCH 4/4] x86/cpu: Use SERIALIZE in sync_core() when available

[peterz]

On Mon, Jul 27, 2020 at 03:05:36PM +0200, peterz@xxxxxxxxxxxxx wrote:
> Yeah, I'm not sure.. the 'funny' thing is that typically call
> sync_core() from an IPI anyway. And the synchronous broadcast IPI is by
> far the most expensive part of that.
> 
> Something like this...
> 
> diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
> index 20e07feb4064..528e049ee1d9 100644
> --- a/arch/x86/kernel/alternative.c
> +++ b/arch/x86/kernel/alternative.c
> @@ -989,12 +989,13 @@ void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
>  
>  static void do_sync_core(void *info)
>  {
> -	sync_core();
> +	/* IRET implies sync_core() */
>  }
>  
>  void text_poke_sync(void)
>  {
>  	on_each_cpu(do_sync_core, NULL, 1);
> +	sync_core();
>  }
>  
>  struct text_poke_loc {

So 'people' have wanted to optimize this for NOHZ_FULL and I suppose
virt as well.

IFF VMENTER is serializing, I suppose we can simply do something like:

bool text_poke_cond(int cpu, void *info)
{
	/*
	 * If we observe the vCPU is preempted, it will do VMENTER
	 * no point in sending an IPI to SERIALIZE.
	 */
	return !vcpu_is_preempted(cpu);
}

void text_poke_sync(void)
{
	smp_call_function_many_cond(cpu_possible_mask,
				    do_sync_core, NULL, 1, text_poke_cond);
	sync_core();
}

The 'same' for NOHZ_FULL, except we need to cmpxchg a value such that
if the cmpxchg() succeeds we know the CPU is in userspace and will
SERIALIZE on the next entry. Much like kvm_flush_tlb_others().


Anyway, that's all hand-wavey.. I'll let someone that cares about those
things write actual patches :-)