Re: [patch 3/5] mm: vmstat: optionally flush per-CPU vmstat counters on return to userspace

[Marcelo Tosatti <mtosatti@redhat.com>]

On Mon, Jul 12, 2021 at 11:05:58AM +0200, Christoph Lameter wrote:
> On Fri, 9 Jul 2021, Marcelo Tosatti wrote:
> 
> > +
> > +	if (!static_branch_unlikely(&vmstat_sync_enabled))
> > +		return;
> > +
> > +	cpu = smp_processor_id();
> > +
> > +	if (housekeeping_cpu(cpu, HK_FLAG_QUIESCE_URET))
> > +		return;
> > +
> > +	per_cpu(vmstat_dirty, smp_processor_id()) = true;
> > +}
> 
> And you are going to insert this into all the performance critical VM
> statistics handling. Inline?

Yes, this is what the patch below is supposed to do (maybe it missed
some statistics?).

The alternative would be some equivalent of need_update on return to
userspace (for all system call returns) (when the HK_FLAG_QUIESCE_URET 
flag is enabled).

> And why do you need to do such things as to determine the processor? At
> mininum do this using this cpu operations like the vmstat functions
> currently do.

OK, will do that and resend.

> And, lucky us, now we also have
> more issues why we should disable preemption etc etc while handling vm
> counters.