mm: hugetlb: fix a race between freeing and dissolving the page

There is a race condition between __free_huge_page()
and dissolve_free_huge_page().

  CPU0:                         CPU1:

  // page_count(page) == 1
  put_page(page)
    __free_huge_page(page)
                                dissolve_free_huge_page(page)
                                  spin_lock(&hugetlb_lock)
                                  // PageHuge(page) && !page_count(page)
                                  update_and_free_page(page)
                                  // page is freed to the buddy
                                  spin_unlock(&hugetlb_lock)
      spin_lock(&hugetlb_lock)
      clear_page_huge_active(page)
      enqueue_huge_page(page)
      // It is wrong, the page is already freed
      spin_unlock(&hugetlb_lock)

The race window is between put_page() and dissolve_free_huge_page().

We should make sure that the page is already on the free list when it is
dissolved.

As a result __free_huge_page would corrupt page(s) already in the buddy
allocator.

Link: https://lkml.kernel.org/r/20210115124942.46403-4-songmuchun@bytedance.com
Fixes: c8721bbb ("mm: memory-hotplug: enable memory hotplug to handle hugepage")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/hugetlb.c

@@ -79,6 +79,21 @@ DEFINE_SPINLOCK(hugetlb_lock);
 static int num_fault_mutexes;
 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
 
+static inline bool PageHugeFreed(struct page *head)
+{
+	return page_private(head + 4) == -1UL;
+}
+
+static inline void SetPageHugeFreed(struct page *head)
+{
+	set_page_private(head + 4, -1UL);
+}
+
+static inline void ClearPageHugeFreed(struct page *head)
+{
+	set_page_private(head + 4, 0);
+}
+
 /* Forward declaration */
 static int hugetlb_acct_memory(struct hstate *h, long delta);
 
@@ -1028,6 +1043,7 @@ static void enqueue_huge_page(struct hstate *h, struct page *page)
 	list_move(&page->lru, &h->hugepage_freelists[nid]);
 	h->free_huge_pages++;
 	h->free_huge_pages_node[nid]++;
+	SetPageHugeFreed(page);
 }
 
 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
@@ -1044,6 +1060,7 @@ static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
 
 		list_move(&page->lru, &h->hugepage_activelist);
 		set_page_refcounted(page);
+		ClearPageHugeFreed(page);
 		h->free_huge_pages--;
 		h->free_huge_pages_node[nid]--;
 		return page;
@@ -1505,6 +1522,7 @@ static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 	spin_lock(&hugetlb_lock);
 	h->nr_huge_pages++;
 	h->nr_huge_pages_node[nid]++;
+	ClearPageHugeFreed(page);
 	spin_unlock(&hugetlb_lock);
 }
 
@@ -1755,6 +1773,7 @@ int dissolve_free_huge_page(struct page *page)
 {
 	int rc = -EBUSY;
 
+retry:
 	/* Not to disrupt normal path by vainly holding hugetlb_lock */
 	if (!PageHuge(page))
 		return 0;
@@ -1771,6 +1790,26 @@ int dissolve_free_huge_page(struct page *page)
 		int nid = page_to_nid(head);
 		if (h->free_huge_pages - h->resv_huge_pages == 0)
 			goto out;
+
+		/*
+		 * We should make sure that the page is already on the free list
+		 * when it is dissolved.
+		 */
+		if (unlikely(!PageHugeFreed(head))) {
+			spin_unlock(&hugetlb_lock);
+			cond_resched();
+
+			/*
+			 * Theoretically, we should return -EBUSY when we
+			 * encounter this race. In fact, we have a chance
+			 * to successfully dissolve the page if we do a
+			 * retry. Because the race window is quite small.
+			 * If we seize this opportunity, it is an optimization
+			 * for increasing the success rate of dissolving page.
+			 */
+			goto retry;
+		}
+
 		/*
 		 * Move PageHWPoison flag from head page to the raw error page,
 		 * which makes any subpages rather than the error page reusable.