On Wed, Nov 09, 2022 at 04:40:26PM -0300, Marcelo Tosatti wrote:
> On Mon, Oct 24, 2022 at 01:03:11PM +0200, Frederic Weisbecker wrote:
> > Lines: 94
> >
> > On Sat, Sep 24, 2022 at 04:22:25PM +0100, Aaron Tomlin wrote:
> > > From: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
> > >
> > > From the vmstat shepherd, for CPUs that have the tick stopped, do not
> > > queue local work to flush the per-CPU vmstats, since in that case the
> > > flush is performed on return to userspace or when entering idle. Also
> > > cancel any delayed work on the local CPU, when entering idle on nohz
> > > full CPUs. Per-CPU pages can be freed remotely from housekeeping CPUs.
> > >
> > > Signed-off-by: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
> > > ---
> > > mm/vmstat.c | 18 +++++++++++++-----
> > > 1 file changed, 13 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/mm/vmstat.c b/mm/vmstat.c
> > > index 472175642bd9..3b9a497965b4 100644
> > > --- a/mm/vmstat.c
> > > +++ b/mm/vmstat.c
> > > @@ -29,6 +29,7 @@
> > > #include <linux/page_ext.h>
> > > #include <linux/page_owner.h>
> > > #include <linux/migrate.h>
> > > +#include <linux/tick.h>
> > >
> > > #include "internal.h"
> > >
> > > @@ -1990,19 +1991,23 @@ static void vmstat_update(struct work_struct *w)
> > > */
> > > void quiet_vmstat(void)
> > > {
> > > + struct delayed_work *dw;
> > > +
> > > if (system_state != SYSTEM_RUNNING)
> > > return;
> > >
> > > if (!is_vmstat_dirty())
> > > return;
> > >
> > > + refresh_cpu_vm_stats(false);
> > > +
> > > /*
> > > - * Just refresh counters and do not care about the pending delayed
> > > - * vmstat_update. It doesn't fire that often to matter and canceling
> > > - * it would be too expensive from this path.
> > > - * vmstat_shepherd will take care about that for us.
> > > + * If the tick is stopped, cancel any delayed work to avoid
> > > + * interruptions to this CPU in the future.
> > > */
> > > - refresh_cpu_vm_stats(false);
> > > + dw = &per_cpu(vmstat_work, smp_processor_id());
> > > + if (delayed_work_pending(dw) && tick_nohz_tick_stopped())
> > > + cancel_delayed_work(dw);
> > > }
> > >
> > > /*
> > > @@ -2024,6 +2029,9 @@ static void vmstat_shepherd(struct work_struct *w)
> > > for_each_online_cpu(cpu) {
> > > struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
> > >
> > > + if (tick_nohz_tick_stopped_cpu(cpu))
> > > + continue;
> > > +
> > > if (!delayed_work_pending(dw) && per_cpu(vmstat_dirty, cpu))
> > > queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
> >
> > All these checks are racy though. You may well eventually:
> >
> > 1) Arm the timer after the CPU has entered in userspace
> > 2) Not arm the timer when the CPU has entered the kernel
> >
> > How about converting that to an IPI instead? This should be a good candidate
> > for the future IPI deferment.
> >
> > Another possible way to go is this:
> >
> > 1) vmstat_shepherd completely ignores nohz_full CPUs
> > 2) vmstat_work is only ever armed locally
> > 3) A nohz_full CPU turning its local vmstat as dirty checks if vmstat_work is
> > pending. If not, queue it, possibly through a self IPI (IRQ_WORK) to get
> > away with current locking context.
>
> I'm afraid there might be workloads where local vmstat touch is a
> hot-path.
>
> > 3) Fold on idle if dirty
> > 4) Fold on user enter and disarm vmstat_work if pending
> >
> > Does that sound possible?
> >
> > Thanks.
>
> I guess so, but proper barriers would also work.
>
> Do you have any particular reason for the 1-4 sequence above
> instead of barriers?
I think a per-CPU atomic variable might be necessary, not just barriers.
Thanks.
