Re: [PATCH v5 1/2] mm,hwpoison: fix race with hugetlb page allocation

From: HORIGUCHI NAOYA(堀口 直也)
Date: Thu May 20 2021 - 03:17:22 EST


On Wed, May 19, 2021 at 03:32:17PM -0700, Mike Kravetz wrote:
> On 5/18/21 4:12 PM, Naoya Horiguchi wrote:
> > From: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>
> >
> > When hugetlb page fault (under overcommitting situation) and
> > memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:
> >
> > CPU0: CPU1:
> >
> > gather_surplus_pages()
> > page = alloc_surplus_huge_page()
> > memory_failure_hugetlb()
> > get_hwpoison_page(page)
> > __get_hwpoison_page(page)
> > get_page_unless_zero(page)
> > zero = put_page_testzero(page)
> > VM_BUG_ON_PAGE(!zero, page)
> > enqueue_huge_page(h, page)
> > put_page(page)
> >
> > __get_hwpoison_page() only checks the page refcount before taking an
> > additional one for memory error handling, which is wrong because there's
> > a time window where compound pages have non-zero refcount during
> > initialization. So make __get_hwpoison_page() check page status a bit
> > more for hugetlb pages.
> >
> > Fixes: ead07f6a867b ("mm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling")
> > Signed-off-by: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>
> > Reported-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> > Cc: stable@xxxxxxxxxxxxxxx # 5.12+
> > ---
> > include/linux/hugetlb.h | 6 ++++++
> > mm/hugetlb.c | 15 +++++++++++++++
> > mm/memory-failure.c | 8 +++++++-
> > 3 files changed, 28 insertions(+), 1 deletion(-)
> >
> > diff --git v5.13-rc2/include/linux/hugetlb.h v5.13-rc2_patched/include/linux/hugetlb.h
> > index b92f25ccef58..790ae618548d 100644
> > --- v5.13-rc2/include/linux/hugetlb.h
> > +++ v5.13-rc2_patched/include/linux/hugetlb.h
> > @@ -149,6 +149,7 @@ bool hugetlb_reserve_pages(struct inode *inode, long from, long to,
> > long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
> > long freed);
> > bool isolate_huge_page(struct page *page, struct list_head *list);
> > +int get_hwpoison_huge_page(struct page *page, bool *hugetlb);
> > void putback_active_hugepage(struct page *page);
> > void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason);
> > void free_huge_page(struct page *page);
> > @@ -339,6 +340,11 @@ static inline bool isolate_huge_page(struct page *page, struct list_head *list)
> > return false;
> > }
> >
> > +static inline int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
> > +{
> > + return 0;
> > +}
> > +
> > static inline void putback_active_hugepage(struct page *page)
> > {
> > }
> > diff --git v5.13-rc2/mm/hugetlb.c v5.13-rc2_patched/mm/hugetlb.c
> > index 95918f410c0f..f138bae3e302 100644
> > --- v5.13-rc2/mm/hugetlb.c
> > +++ v5.13-rc2_patched/mm/hugetlb.c
> > @@ -5847,6 +5847,21 @@ bool isolate_huge_page(struct page *page, struct list_head *list)
> > return ret;
> > }
> >
> > +int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
> > +{
> > + int ret = 0;
> > +
> > + *hugetlb = false;
> > + spin_lock_irq(&hugetlb_lock);
> > + if (PageHeadHuge(page)) {
> > + *hugetlb = true;
> > + if (HPageFreed(page) || HPageMigratable(page))
> > + ret = get_page_unless_zero(page);
> > + }
> > + spin_unlock_irq(&hugetlb_lock);
> > + return ret;
> > +}
> > +
> > void putback_active_hugepage(struct page *page)
> > {
> > spin_lock_irq(&hugetlb_lock);
> > diff --git v5.13-rc2/mm/memory-failure.c v5.13-rc2_patched/mm/memory-failure.c
> > index 85ad98c00fd9..353c6177e489 100644
> > --- v5.13-rc2/mm/memory-failure.c
> > +++ v5.13-rc2_patched/mm/memory-failure.c
> > @@ -959,8 +959,14 @@ static int page_action(struct page_state *ps, struct page *p,
> > static int __get_hwpoison_page(struct page *page)
> > {
> > struct page *head = compound_head(page);
> > + int ret = 0;
> > + bool hugetlb = false;
> > +
> > + ret = get_hwpoison_huge_page(head, &hugetlb);
> > + if (hugetlb)
> > + return ret;
> >
>
> Hello Naoya,
>
> Thanks for your continued efforts. However, I believe the race still
> exists. Unless I am mistaken, it is possible that page is in the hugetlb
> allocation patch and racing with __get_hwpoison_page() as follows:

Hi Mike, thank you for the investigation.

>
> CPU0: CPU1:
>
> gather_surplus_pages()
> page = alloc_surplus_huge_page()
> page = alloc_fresh_huge_page()
> page = alloc_buddy_huge_page()
> memory_failure_hugetlb()
> get_hwpoison_page(page)
> __get_hwpoison_page(page)
> get_hwpoison_huge_page()
> /* Note that PageHuge()
> is false, so hugetlb
> not set */
> PageTransHuge(head) false

It seems that PageTransHuge returns true in this race condition because it
simply checks PG_head flag. But anyway, get_page_unless_zero() is called for
a hugetlb in this race, which is problematic.

> prep_new_huge_page(page)
> /* Now PageHuge() becomes true */
> get_page_unless_zero(page)
>
> I am not sure if it is possible to handle this race in the memory error
> code.

I think that __get_hwpoison_page() might not properly call get_page_unless_zero().
Looking at other callers of this function, most(*) of them are calling it in
the context where a given page is pinned and in-use, and get_page_unless_zero()
is used to detect the race with page freeing. In the current version,
__get_hwpoison_page() calls get_page_unless_zero() without caring for such an
assumption, which might be the root cause of the race with hugetlb page allocation.

# (*) It seems to me that do_migrate_range() might have the same issue
# around get_page_unless_zero().

So I think of inserting the check to comply with the assumption of
get_hwpoison_huge_page() like below:

ret = get_hwpoison_huge_page(head, &hugetlb);
if (hugetlb)
return ret;

if (!PageLRU(head) && !__PageMovable(head))
return 0;

if (PageTransHuge(head)) {
...
}

if (get_page_unless_zero(head)) {
...
}

return 0;

The newly added checks should work to prevent the above race, then get_any_page()
should retry and grab the page properly as a stable hugetlb page.

> I can not think of a way to avoid potentially incrementing the
> ref count on a hugetlb page as it is being created. There is nothing
> synchronizing this in the hugetlb code.
>
> When Muchun first proposed a fix to the race, the idea was to catch the
> race in the hugetlb code. Michal suggested that the memory error code
> be more careful in modifying ref counts. I would wait a bit to see if
> someone has a good idea how this can be done. We 'may' need to revisit
> the approach suggested by Muchun.

If the above approach is still broken, let's revisit Muchun's approach.

Thanks,
Naoya Horiguchi