Re: [PATCH 37/49] x86/alternatives: Move text_poke_array completion from smp_text_poke_batch_finish() and smp_text_poke_batch_flush() to smp_text_poke_batch_process()

[Nikolay Borisov]

On 3.04.25 г. 18:29 ч., Ingo Molnar wrote:
> * Nikolay Borisov <nik.borisov@xxxxxxxx> wrote:
>
> > I meant doing this:
> >
> > diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
> > index 5b1a6252a4b9..b6a781b9de26 100644
> > --- a/arch/x86/kernel/alternative.c
> > +++ b/arch/x86/kernel/alternative.c
> > @@ -2587,12 +2587,16 @@ noinstr int smp_text_poke_int3_trap_handler(struct
> > pt_regs *regs)
> >    *               replacing opcode
> >    *     - SMP sync all CPUs
> >    */
> > -static void smp_text_poke_batch_process(void)
> > +void smp_text_poke_batch_finish(void)
> >   {
> >          unsigned char int3 = INT3_INSN_OPCODE;
> >          unsigned int i;
> >          int do_sync;
> >
> > +
> > +       if (!text_poke_array.nr_entries)
> > +               return;
>
> > -               smp_text_poke_batch_process();
> > +               smp_text_poke_batch_finish();
>
> I suppose we could do this - it adds one more check to
> smp_text_poke_batch_add() though.

poke_batch_finish you meant?

> > AFAICS this doesn't change the semantics. I.e smp_text_poke_batch_add
> > will call poke_batch_finish iff the address to be added violates the
> > sorted order of text_poke_array. The net effect is we have 1 less
> > function name to care about.
>
> Yeah, it doesn't change semantics, but it's a very small
> deoptimization.
> > Mind sending a patch? It does simplify the facility some more and that
> single branch will wash away against costs like the CR3 flushes done

Given that poke_batch_finish does a cond_resched and sync_each_cpu which 
is an IPI can it even be considered a performance critical path ?

> ...
>
> Thanks,
>
> 	Ingo