On Tue, Jun 01, 2021 at 11:47:13AM +0200, Michal Hocko wrote:

While you are touching this and want to drill all the way down then it
would be reasonable to drop pgdat resize locks as well.
It is only used in the early boot code and we have one executing thread
context per numa node during the deferred initialization. I haven't
checked all potential side effects the lock might have but it sounds
like there is quite some clean up potential over there.

I am not sure about that. True is that deferred_init_memmap() gets executed
on numa-thread so it's not a problem for itself, but we also have deferred_grow_zone().
It might be that while deferred_init_memmap() is running, we also have calls to
deferred_grow_zone() for the same node and that would cause some trouble wrt.
first_deferred_pfn. I need to double check it, but IIRC, that is the case.


-- 
Oscar Salvador
SUSE L3