Re: [PATCH v1] mm, hwpoison: fix condition in free hugetlb page path

[Naoya Horiguchi <naoya.horiguchi@linux.dev>]

On Fri, Dec 10, 2021 at 08:02:08PM +0900, Naoya Horiguchi wrote:
> From: Naoya Horiguchi <naoya.horiguchi@nec.com>
> 
> When a memory error hits a tail page of a free hugepage,
> __page_handle_poison() is expected to be called to isolate the error in
> 4kB unit, but it's not called due to the outdated if-condition in
> memory_failure_hugetlb().  This loses the chance to isolate the error in
> the finer unit, so it's not optimal.  Drop the condition.
> 
> This "(p != head && TestSetPageHWPoison(head)" condition is based on the
> old semantics of PageHWPoison on hugepage (where PG_hwpoison flag was
> set on the subpage), so it's not necessray any more.  By getting to set
> PG_hwpoison on head page for hugepages, concurrent error events on
> different subpages in a single hugepage can be prevented by
> TestSetPageHWPoison(head) at the beginning of memory_failure_hugetlb().
> So dropping the condition should not reopen the race window originally
> mentioned in commit b985194c8c0a ("hwpoison, hugetlb:
> lock_page/unlock_page does not match for handling a free hugepage")
> 
> Reported-by: Fei Luo <luofei@unicloud.com>
> Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
> Cc: <stable@vger.kernel.org> # v5.14+
> ---
> I set v5.14+ for stable trees because the base code was greatly changed
> by commit 0ed950d1f281 ("mm,hwpoison: make get_hwpoison_page() call
> get_any_page()"), and this patch is not cleanly applicable, although the
> original issue was introduced more previously.
> ---
>  mm/memory-failure.c | 21 +++++++--------------
>  1 file changed, 7 insertions(+), 14 deletions(-)
> 
> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> index 8f0ee5b08696..68d9a35f8908 100644
> --- a/mm/memory-failure.c
> +++ b/mm/memory-failure.c
> @@ -1521,24 +1521,17 @@ static int memory_failure_hugetlb(unsigned long pfn, int flags)
>  	if (!(flags & MF_COUNT_INCREASED)) {
>  		res = get_hwpoison_page(p, flags);
>  		if (!res) {
> -			/*
> -			 * Check "filter hit" and "race with other subpage."
> -			 */
>  			lock_page(head);
> -			if (PageHWPoison(head)) {
> -				if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
> -				    || (p != head && TestSetPageHWPoison(head))) {
> +			if (hwpoison_filter(p)) {
> +				if (TestClearPageHWPoison(head))
>  					num_poisoned_pages_dec();
> -					unlock_page(head);
> -					return 0;
> -				}
> +				unlock_page(head);
> +				return 0;
>  			}
>  			unlock_page(head);
> -			res = MF_FAILED;
> -			if (__page_handle_poison(p)) {
> -				page_ref_inc(p);
> -				res = MF_RECOVERED;
> -			}
> +			res = MF_RECOVERED;
> +			if (!page_handle_poison(p, true, false))
> +				res = MF_FAILED;

Sorry, I just found that this change broke "HardwareCorrupted" counter
because page_handle_poison() calls num_poisoned_pages_inc() so the counter
was incremented twice in total by a single error event.  page_handle_poison()
is supposed to be called only in soft-offline context, and __page_handle_poison()
should be called in hard-offline context.

Andrew, could you fold the following diff on
mm-hwpoison-fix-condition-in-free-hugetlb-page-path.patch in your tree?
(Or if desirable, I'll send a updated full patch.)

diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index 68d9a35f8908..ee51d6410f9a 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -1529,9 +1529,11 @@ static int memory_failure_hugetlb(unsigned long pfn, int flags)
 				return 0;
 			}
 			unlock_page(head);
-			res = MF_RECOVERED;
-			if (!page_handle_poison(p, true, false))
-				res = MF_FAILED;
+			res = MF_FAILED;
+			if (__page_handle_poison(p)) {
+				page_ref_inc(p);
+				res = MF_RECOVERED;
+			}
 			action_result(pfn, MF_MSG_FREE_HUGE, res);
 			return res == MF_RECOVERED ? 0 : -EBUSY;
 		} else if (res < 0) {


Thanks,
Naoya Horiguchi