Re: [PATCH v4 4/6] mm/memory-hotplug: add finite retries in offline_pages() if migration fails

[Joanne Koong <joannelkoong@gmail.com>]

On Fri, Nov 8, 2024 at 2:16 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
>
> On Fri, Nov 08, 2024 at 01:42:15PM -0800, Joanne Koong wrote:
> > On Fri, Nov 8, 2024 at 1:27 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> > >
> > > On Fri, Nov 08, 2024 at 08:00:25PM +0100, David Hildenbrand wrote:
> > > > On 08.11.24 19:56, David Hildenbrand wrote:
> > > > > On 08.11.24 18:33, SeongJae Park wrote:
> > > > > > + David Hildenbrand
> > > > > >
> > > > > > On Thu, 7 Nov 2024 15:56:12 -0800 Joanne Koong <joannelkoong@gmail.com> wrote:
> > > > > >
> > > > > > > In offline_pages(), do_migrate_range() may potentially retry forever if
> > > > > > > the migration fails. Add a return value for do_migrate_range(), and
> > > > > > > allow offline_page() to try migrating pages 5 times before erroring
> > > > > > > out, similar to how migration failures in __alloc_contig_migrate_range()
> > > > > > > is handled.
> > > > > >
> > > > > > I'm curious if this could cause unexpected behavioral differences to memory
> > > > > > hotplugging users, and how '5' is chosen.  Could you please enlighten me?
> > > > > >
> > > > >
> > > > > I'm wondering how much more often I'll have to nack such a patch. :)
> > > >
> > > > A more recent discussion: https://lore.kernel.org/linux-mm/52161997-15aa-4093-a573-3bfd8da14ff1@fujitsu.com/T/#mdda39b2956a11c46f8da8796f9612ac007edbdab
> > > >
> > > > Long story short: this is expected and documented
> > >
> > > Thanks David for the background.
> > >
> > > Joanne, simply drop this patch. It is not required for your series.
> >
> > Awesome, I'm happy to drop this patch.
> >
> > Just curious though - don't we need this in order to mitigate the
> > scenario where if an unprivileged fuse server never completes
> > writeback, we don't run into this infinite loop? Or is it that memory
> > hotplugging is always initiated from userspace so if it does run into
> > an infinite loop (like also in that thread David linked), userspace is
> > responsible for sending a signal to terminate it?
>
> I think irrespective of the source of the hotplug, the current behavior
> of infinite retries in some cases is documented and kind of expected, so
> no need to fix it. (Though I don't know all the source of hotplug.)
>

Awesome, this sounds great. Thanks!
>