On 1/6/21 12:02 PM, Michal Hocko wrote: > On Wed 06-01-21 11:30:25, Mike Kravetz wrote: >> On 1/6/21 8:35 AM, Michal Hocko wrote: >>> On Wed 06-01-21 16:47:35, Muchun Song wrote: >>>> Because we only can isolate a active page via isolate_huge_page() >>>> and hugetlbfs_fallocate() forget to mark it as active, we cannot >>>> isolate and migrate those pages. >>> >>> I've little bit hard time to understand this initially and had to dive >>> into the code to make sense of it. I would consider the following >>> wording easier to grasp. Feel free to reuse if you like. >>> " >>> If a new hugetlb page is allocated during fallocate it will not be >>> marked as active (set_page_huge_active) which will result in a later >>> isolate_huge_page failure when the page migration code would like to >>> move that page. Such a failure would be unexpected and wrong. >>> " >>> >>> Now to the fix. I believe that this patch shows that the >>> set_page_huge_active is just too subtle. Is there any reason why we >>> cannot make all freshly allocated huge pages active by default? >> >> I looked into that yesterday. The primary issue is in page fault code, >> hugetlb_no_page is an example. If page_huge_active is set, then it can >> be isolated for migration. So, migration could race with the page fault >> and the page could be migrated before being added to the page table of >> the faulting task. This was an issue when hugetlb_no_page set_page_huge_active >> right after allocating and clearing the huge page. Commit cb6acd01e2e4 >> moved the set_page_huge_active after adding the page to the page table >> to address this issue. > > Thanks for the clarification. I was not aware of this subtlety. The > existing comment is not helping much TBH. I am still digesting the > suggested race. The page is new and exclusive and not visible via page > tables yet, so the only source of the migration would be pfn based > (hotplug, poisoning), right? That is correct. > Btw. s@set_page_huge_active@set_page_huge_migrateable@ would help > readability IMHO. With a comment explaining that this _has_ to be called > after the page is fully initialized. Agree, I will add that as a future enhancement. -- Mike Kravetz
