On Wed, Apr 01, 2020 at 06:00:48PM +0200, Michal Hocko wrote:
> On Wed 01-04-20 17:50:22, David Hildenbrand wrote:
> > On 01.04.20 17:42, Michal Hocko wrote:
> > > I am sorry but I have completely missed this patch.
> > >
> > > On Wed 11-03-20 20:38:48, Shile Zhang wrote:
> > >> When 'CONFIG_DEFERRED_STRUCT_PAGE_INIT' is set, 'pgdatinit' kthread will
> > >> initialise the deferred pages with local interrupts disabled. It is
> > >> introduced by commit 3a2d7fa8a3d5 ("mm: disable interrupts while
> > >> initializing deferred pages").
> > >>
> > >> On machine with NCPUS <= 2, the 'pgdatinit' kthread could be bound to
> > >> the boot CPU, which could caused the tick timer long time stall, system
> > >> jiffies not be updated in time.
> > >>
> > >> The dmesg shown that:
> > >>
> > >> [ 0.197975] node 0 initialised, 32170688 pages in 1ms
> > >>
> > >> Obviously, 1ms is unreasonable.
> > >>
> > >> Now, fix it by restore in the pending interrupts for every 32*1204 pages
> > >> (128MB) initialized, give the chance to update the systemd jiffies.
> > >> The reasonable demsg shown likes:
> > >>
> > >> [ 1.069306] node 0 initialised, 32203456 pages in 894ms
> > >>
> > >> Fixes: 3a2d7fa8a3d5 ("mm: disable interrupts while initializing deferred pages").
> > >
> > > I dislike this solution TBH. It effectivelly conserves the current code
> > > and just works around the problem. Why do we hold the IRQ lock here in
> > > the first place? This is an early init code and a very limited code is
> > > running at this stage. Certainly nothing memory hotplug related which
> > > should be the only path really interested in the resize lock AFAIR.
> >
> > Yeah, I don't think ACPI and friends are up yet.
>
> Just to save somebody time to check. The deferred initialization blocks
> the further boot until all workders are done - see page_alloc_init_late
> (kernel_init path).
Ha, I just finished following all the hotplug paths to check this out, and as
you all know there are a *lot* :-) Well at least we're in agreement.
> > > This needs a double checking but I strongly believe that the lock can be
> > > simply dropped in this path.
This is what my fix does, it limits the time the resize lock is held.
