On Wed, Mar 16, 2022 at 03:51:35PM -0700, Mike Kravetz wrote: > On 3/16/22 05:07, Naoya Horiguchi wrote: > > From: Miaohe Lin <linmiaohe@xxxxxxxxxx> > > > > There is a race condition between memory_failure_hugetlb() and hugetlb > > free/demotion, which causes setting PageHWPoison flag on the wrong page. > > The one simple result is that wrong processes can be killed, but another > > (more serious) one is that the actual error is left unhandled, so no one > > prevents later access to it, and that might lead to more serious results > > like consuming corrupted data. > > > > Think about the below race window: > > > > CPU 1 CPU 2 > > memory_failure_hugetlb > > struct page *head = compound_head(p); > > hugetlb page might be freed to > > buddy, or even changed to another > > compound page. > > > > get_hwpoison_page -- page is not what we want now... > > > > The compound_head is called outside hugetlb_lock, so the head is not > > reliable. > > > > So set PageHWPoison flag after passing prechecks. And to detect > > potential violation, this patch also introduces a new action type > > MF_MSG_DIFFERENT_PAGE_SIZE. > > Thanks for squashing these patches. > > In my testing, there is a change in behavior that may not be intended. > > My test strategy is: > - allocate two hugetlb pages > - create a mapping which reserves those two pages, but does not fault them in > - as a result, the pages are on the free list but can not be freed > - inject error on a subpage of one of the huge pages > - echo 0xYYY > /sys/kernel/debug/hwpoison/corrupt-pfn > - memory error code will call dissolve_free_huge_page > - dissolve_free_huge_page returns -EBUSY because > h->free_huge_pages - h->resv_huge_pages == 0 > - We never end up setting Poison on the page with error or head page > - Huge page sitting on free list with error in subpage and not marked > - huge page with error could be given to an application or returned to buddy > > Prior to this change, Poison would be set on the head page You're right, this is not intended. I should've kept current behavior for this case (set PageHWPoison flag on the head page, and no refcnt taken), so I'll update the patch. In this case the hwpoisoned hugepage remains in free list, but dequeue_huge_page_node_exact() checks the PageHWPoison flag not to be allocated, that's OK. It might not be optimal that the hwpoisoned free hugepage is in the list for long, so some mechanism to handle it in a delayed manner. One way is to call dissolve_free_huge_page() when a hwpoisoned page is found in dequeue_huge_page_node_exact(). If the dissolving succeeds, it's OK. If it fails, keep it as-is expecting to be dissolved next time. Another related problem is that when hwpoison hugepage is listed in free list, the information about raw error offset is lost. So if hugepage pool is shrinked and the hwpoison hugepage is freed to buddy, the PageHWPoison flag remains on the ex-head page. So we need somehow keep error offset. Anyway, this will be done in separate work, I'll just fix the problem you mentioned. Thank you very much, Naoya Horiguchi > > I do not think this was an intended change in behavior. But, perhaps it is > all we can do in this case? Sorry for not being able to look more closely > at the code right now. > -- > Mike Kravetz
