Re: [PATCH] mm/memory_hotplug: drain per-cpu pages again during memory offline

[Pavel Tatashin]

> Another alternative would be to enable/disable static branch only from
> users who really care but this is quite tricky because how do you tell
> you need or not? It seems that alloc_contig_range would be just fine
> with a weaker semantic because it would "only" to a spurious failure.
> Memory hotplug on the other hand really needs to have a point where
> nobody interferes with the offlined memory so it could ask for a
> stronger semantic.
>
> Yet another option would be to make draining stronger and actually
> guarantee there are no in-flight pages to be freed to the pcp list.
> One way would be to tweak pcp->high and implement a strong barrier
> (IPI?) to sync with all CPUs. Quite expensive, especially when there are
> many draining requests (read cma users because hotplug doesn't really
> matter much as it happens seldom).
>
> So no nice&cheap solution I can think of...

I think start_isolate_page_range() should not be doing page draining
at all. It should isolate ranges, meaning set appropriate flags, but
draining should be performed by the users when appropriate: next to
lru_add_drain_all() calls both in CMA and hotplug.
Currently, the way start_isolate_page_range() drains pages is very
inefficient. It calls drain_all_pages() for every valid page block,
which is a slow call as it starts a thread per cpu, and waits for
those threads to finish before returning.

We could optimize by moving the drain_all_pages() calls from
set_migratetype_isolate() to start_isolate_page_range() and call it
once for every different zone, but both current users of this
interface guarantee that all pfns [start_pfn, end_pfn] are within the
same zone, and I think we should keep it this way, so again the extra
traversal is going to be overhead overhead.

This way we will have on average only a single drain per hot-remove
(instead of one per block), and also it is going to be symmetric only
in one place. Faster hot-remove and cma alloc, and no race, imo
win-win.

Pasha