Re: [PATCH 1/4] mm/memory-failure.c: fix race with changing page compound again

[HORIGUCHI NAOYA(堀口 直也)]

On Mon, Mar 07, 2022 at 11:44:20AM +0800, Miaohe Lin wrote:
> On 2022/3/5 3:32, Mike Kravetz wrote:
> > On 3/4/22 00:26, HORIGUCHI NAOYA(堀口 直也) wrote:
> >> On Mon, Feb 28, 2022 at 10:02:42PM +0800, Miaohe Lin wrote:
> >>> There is a race window where we got the compound_head, the hugetlb page
> >>> could be freed to buddy, or even changed to another compound page just
> >>> before we try to get hwpoison page. If this happens, just bail out.
> >>
> >> I think that when some hugetlb page is about to change into other type/size
> >> of compound page, it has to go through buddy allocator because hugetlb pages
> >> are maintained in separate memory allocator and they never change into other
> >> normal state directly.  memory_failure_hugetlb() takes refcount before
> >> lock_page(), so the hugetlb page seems not change between get_hwpoison_page()
> >> and lock_page(). So it this new check really necessary?
> > 
> > A hugetlb page could change size without going through buddy via the new
> > demote functionality [1].  Only hugetlb pages on the hugetlb free list can
> > be demoted.  
> > 
> > We should not demote a page if poison is set.  However, there is no check in
> > the demote code.  IIUC, poison is set early in the memory error handling
> > process, even before taking ref on page.  Demote code needs to be fixed so
> > that poisoned pages are not demoted.  I can do that.
> > 
> > With this change in place, then I think Naoya's statement that hugetlb pages
> > can not change state is correct and this patch is not necessary.
> > 
> 
> Sorry for my confusing commit words. What I mean to tell is indeed the below race:
>   CPU 1							CPU 2
>   memory_failure_hugetlb
>   struct page *head = compound_head(p);
> 							hugetlb page is freed to buddy, or
> 							even changed to another compound page
> 							as we haven't held the page refcnt now
>   get_hwpoison_page -- page is not what we want now...
> 
> Does this make sense for both of you? Many thanks for comment and reply! :)

Thanks for elaboration, I agree with you (I simply overlooked this race, sorry).
And please add this in the commit log.

> +
> +       /**
> +        * The page could have changed compound pages due to race window.
> +        * If this happens just bail out.
> +        */
> +       if (!PageHuge(p) || compound_head(p) != head) {
> +               action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
> +               res = -EBUSY;
> +               goto out;
> +       }

Let me have one comment on the diff. The result code MF_MSG_DIFFERENT_COMPOUND
might not fit when PageHuge is false in the check (because it's no longer a
compound page).  Maybe you may invent another result code, or changes
MF_MSG_DIFFERENT_COMPOUND (for example) to MF_MSG_DIFFERENT_PAGE_SIZE?

> 
> > Does that sound reasonable?
> > 
> > [1] https://lore.kernel.org/linux-mm/20211007181918.136982-1-mike.kravetz@xxxxxxxxxx/
> 
> This is really a nice feature. Thanks.

Yes. Hugepage demotion after entering memory_failure_hugetlb() should
cause the race, so we need the check after page lock.

Thanks,
Naoya Horiguchi