Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Peter Xu <peterx@redhat.com>
To:     linux-kernel@vger.kernel.org, linux-mm@kvack.org
Cc:     Andrew Morton <akpm@linux-foundation.org>, peterx@redhat.com,
        Muchun Song <songmuchun@bytedance.com>,
        Mike Kravetz <mike.kravetz@oracle.com>,
        Naoya Horiguchi <naoya.horiguchi@nec.com>
Subject: [PATCH] mm/hugetlb: Fix pgtable lock on pmd sharing
Date:   Mon, 12 Jun 2023 12:04:20 -0400
Message-Id: <20230612160420.809818-1-peterx@redhat.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Huge pmd sharing operates on PUD not PMD, huge_pte_lock() is not suitable
in this case because it should only work for last level pte changes, while
pmd sharing is always one level higher.

Meanwhile, here we're locking over the spte pgtable lock which is even not
a lock for current mm but someone else's.

It seems even racy on operating on the lock, as after put_page() of the
spte pgtable page logically the page can be released, so at least the
spin_unlock() needs to be done after the put_page().

No report I am aware, I'm not even sure whether it'll just work on taking
the spte pmd lock, because while we're holding i_mmap read lock it probably
means the vma interval tree is frozen, all pte allocators over this pud
entry could always find the specific svma and spte page, so maybe they'll
serialize on this spte page lock?  Even so, doesn't seem to be expected.
It just seems to be an accident of cb900f412154.

Fix it with the proper pud lock (which is the mm's page_table_lock).

Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Fixes: cb900f412154 ("mm, hugetlb: convert hugetlbfs to use split pmd lock")
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 mm/hugetlb.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index dfa412d8cb30..270ec0ecd5a1 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -7133,7 +7133,6 @@ pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 	unsigned long saddr;
 	pte_t *spte = NULL;
 	pte_t *pte;
-	spinlock_t *ptl;
 
 	i_mmap_lock_read(mapping);
 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
@@ -7154,7 +7153,7 @@ pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 	if (!spte)
 		goto out;
 
-	ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
+	spin_lock(&mm->page_table_lock);
 	if (pud_none(*pud)) {
 		pud_populate(mm, pud,
 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
@@ -7162,7 +7161,7 @@ pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 	} else {
 		put_page(virt_to_page(spte));
 	}
-	spin_unlock(ptl);
+	spin_unlock(&mm->page_table_lock);
 out:
 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
 	i_mmap_unlock_read(mapping);
-- 
2.40.1