Re: [PATCH v3] mm/page_owner: fix for deferred struct page init

[Michal Hocko]

On Fri 04-01-19 10:01:40, Qian Cai wrote:
> On 1/4/19 8:09 AM, Michal Hocko wrote:
> >> Here is the number without DEFERRED_STRUCT_PAGE_INIT.
> >>
> >> == page_ext_init() after page_alloc_init_late() ==
> >> Node 0, zone DMA: page owner found early allocated 0 pages
> >> Node 0, zone DMA32: page owner found early allocated 7009 pages
> >> Node 0, zone Normal: page owner found early allocated 85827 pages
> >> Node 4, zone Normal: page owner found early allocated 75063 pages
> >>
> >> == page_ext_init() before kmemleak_init() ==
> >> Node 0, zone DMA: page owner found early allocated 0 pages
> >> Node 0, zone DMA32: page owner found early allocated 6654 pages
> >> Node 0, zone Normal: page owner found early allocated 41907 pages
> >> Node 4, zone Normal: page owner found early allocated 41356 pages
> >>
> >> So, it told us that it will miss tens of thousands of early page allocation call
> >> sites.
> > 
> > This is an answer for the first part of the question (how much). The
> > second is _do_we_care_?
> 
> Well, the purpose of this simple "ugly" ifdef is to avoid a regression for the
> existing page_owner users with DEFERRED_STRUCT_PAGE_INIT deselected that would
> start to miss tens of thousands early page allocation call sites.

I am pretty sure we will hear about that when that happens. And act
accordingly.

> The other option I can think of to not hurt your eyes is to rewrite the whole
> page_ext_init(), init_page_owner(), init_debug_guardpage() to use all early
> functions, so it can work in both with DEFERRED_STRUCT_PAGE_INIT=y and without.
> However, I have a hard-time to convince myself it is a sensible thing to do.

Or simply make the page_owner initialization only touch the already
initialized memory. Have you explored that option as well?

Look, I am trying to push for a clean solution. Hacks I have seen so
far are not convincing. You have identified a regression and as such I
would consider the most straightforward to revert the buggy commit. If
you want to improve the situation then I would suggest to think about
something that is more robust than ifdefed hacks. It might be more work
but it also might be a better thing long term.

If you think I am asking too much then you are free to ignore my
opinion. I am not a maintainer of the page_owner code.
-- 
Michal Hocko
SUSE Labs