Re: [PATCH] mm/memory_hotplug: Fix remove_memory() lockdep splat

From: Dan Williams
Date: Fri Jan 24 2020 - 13:05:05 EST


On Fri, Jan 24, 2020 at 4:56 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Fri 10-01-20 13:27:24, Dan Williams wrote:
> > On Fri, Jan 10, 2020 at 9:42 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> [...]
> > > For your reference (roughly 5 months ago, so not that old)
> > >
> > > https://lkml.kernel.org/r/20190724143017.12841-1-david@xxxxxxxxxx
> >
> > Oh, now I see the problem. You need to add that lock so far away from
> > the __add_memory() to avoid lock inversion problems with the
> > acpi_scan_lock. The organization I was envisioning would not work
> > without deeper refactoring.
>
> Sorry to come back to this late. Has this been resolved?

The mem_hotplug_lock lockdep splat fix in this patch has not landed.
David and I have not quite come to consensus on how to resolve online
racing removal. IIUC David wants that invalidation to be
pages_correctly_probed(), I would prefer it to be directly tied to the
object, struct memory_block, that remove_memory_block_devices() has
modified, mem->section_count = 0.

...or are you referring to the discussion about acpi_scan_lock()? I
came around to agreeing with your position that documenting was better
than adding superfluous locking especially because the
acpi_scan_lock() is take so far away from where the device_hotplug
lock is needed.