On 05/24/2017 01:38 PM, Xishi Qiu wrote:
>>
>> Race condition with what? Who else would isolate our pages?
>>
>
> Hi Vlastimil,
>
> I find the root cause, if the page was not cached on the current cpu,
> lru_add_drain() will not push it to LRU. So we should handle fail
> case in mlock_vma_page().
Yeah that would explain it.
> follow_page_pte()
> ...
> if (page->mapping && trylock_page(page)) {
> lru_add_drain(); /* push cached pages to LRU */
> /*
> * Because we lock page here, and migration is
> * blocked by the pte's page reference, and we
> * know the page is still mapped, we don't even
> * need to check for file-cache page truncation.
> */
> mlock_vma_page(page);
> unlock_page(page);
> }
> ...
>
> I think we should add yisheng's patch, also we should add the following
> change.
> I think it is better than use lru_add_drain_all().
I agree about yisheng's fix (but v2 didn't address my comments). I don't
think we should add the hunk below, as that deviates from the rest of
the design.
Thanks,
Vlastimil
> diff --git a/mm/mlock.c b/mm/mlock.c
> index 3d3ee6c..ca2aeb9 100644
> --- a/mm/mlock.c
> +++ b/mm/mlock.c
> @@ -88,6 +88,11 @@ void mlock_vma_page(struct page *page)
> count_vm_event(UNEVICTABLE_PGMLOCKED);
> if (!isolate_lru_page(page))
> putback_lru_page(page);
> + else {
> + ClearPageMlocked(page);
> + mod_zone_page_state(page_zone(page), NR_MLOCK,
> + -hpage_nr_pages(page));
> + }
> }
> }
>
> Thanks,
> Xishi Qiu
>
