On 1/31/22 05:24, Matthew Wilcox wrote:
> On Mon, Jan 31, 2022 at 05:44:35AM +0000, HORIGUCHI NAOYA(堀口 直也) wrote:
>> On Sun, Jan 30, 2022 at 09:14:21PM +0000, Matthew Wilcox wrote:
>>> On Sun, Jan 30, 2022 at 12:58:17PM -0800, David Rientjes wrote:
>>>> On Sun, 30 Jan 2022, Matthew Wilcox (Oracle) wrote:
>>>>
>>>>> Hardware poison is tracked on a per-page basis, not on the head page.
>>>>>
>>>>> Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
>>>>> ---
>>>>> mm/rmap.c | 4 ++--
>>>>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>>> index 6a1e8c7f6213..09b08888120e 100644
>>>>> --- a/mm/rmap.c
>>>>> +++ b/mm/rmap.c
>>>>> @@ -1553,7 +1553,7 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>>>>> /* Update high watermark before we lower rss */
>>>>> update_hiwater_rss(mm);
>>>>>
>>>>> - if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
>>>>> + if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) {
>>>>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
>>>>> if (PageHuge(page)) {
>>>>> hugetlb_count_sub(compound_nr(page), mm);
>>>>> @@ -1873,7 +1873,7 @@ static bool try_to_migrate_one(struct page *page, struct vm_area_struct *vma,
>>>>> * memory are supported.
>>>>> */
>>>>> subpage = page;
>>>>> - } else if (PageHWPoison(page)) {
>>>>> + } else if (PageHWPoison(subpage)) {
>>>>> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
>>>>> if (PageHuge(page)) {
>>>>> hugetlb_count_sub(compound_nr(page), mm);
>>>>
>>>> This looks correct. Correct me if I'm wrong that this is for consistency
>>>> and cleanup and that there is no bug being fixed by this, however.
>>>
>>> Oh, no, I think there's a real bug here. It's just that we're looking
>>> at an uncommon & hence rarely-tested scenario -- a memory fault in the
>>> middle of a THP (in mainline; obviously it'll be a little more common
>>> with arbitrary sized folios). I don't do HWPoison testing myself, so
>>> this was by inspection and not from testing. A scenario where things
>>> would go wrong is a memory error on a non-head-page would go unnoticed
>>> when migrating or unmapping. Contrariwise, if there's a hardware error
>>> on a head page, all the subpages get treated as poisoned, even though
>>> they shouldn't be.
>>
>> Thank you for reporting. As you point out, the current check does not
>> handle thp properly. The reason of checking head page here is to handle
>> hwpoisoned hugetlb (which has PG_hwpoison on the head page even if the error
>> is on any of tail page). So I think that the proper fix is to add a helper
>> function to check page type (normal, thp, or hugetlb) as well as PageHWPoison.
>
> I think this handles HugeTLB pages correctly:
>
> subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
>
> As I understand the HugeTLB code, pvmw.pte refers to the head page, not
> the subpage (unlike a PTE-mapped THP page, where it would refer to the
> precise page). But I cheerfully admit that the intricacies of the
> HugeTLB code are not something I'm an expert on.
Your understanding is correct. Here is the comment for the routine which
sets pvmw.pte.
* For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
* regardless of which page table level the page is mapped at. @pvmw->pmd is
* NULL.
Just another thought. With new hugetlb vmmemmap optimizations, it is not
possible to set poison on hugetlb tail pages until after allocating struct
pages.
--
Mike Kravetz
