mm/huge_memory: fix dereferencing invalid pmd migration entry

When migrating a THP, concurrent access to the PMD migration entry during
a deferred split scan can lead to an invalid address access, as
illustrated below.  To prevent this invalid access, it is necessary to
check the PMD migration entry and return early.  In this context, there is
no need to use pmd_to_swp_entry and pfn_swap_entry_to_page to verify the
equality of the target folio.  Since the PMD migration entry is locked, it
cannot be served as the target.

Mailing list discussion and explanation from Hugh Dickins: "An anon_vma
lookup points to a location which may contain the folio of interest, but
might instead contain another folio: and weeding out those other folios is
precisely what the "folio != pmd_folio((*pmd)" check (and the "risk of
replacing the wrong folio" comment a few lines above it) is for."

BUG: unable to handle page fault for address: ffffea60001db008
CPU: 0 UID: 0 PID: 2199114 Comm: tee Not tainted 6.14.0+ #4 NONE
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
RIP: 0010:split_huge_pmd_locked+0x3b5/0x2b60
Call Trace:
<TASK>
try_to_migrate_one+0x28c/0x3730
rmap_walk_anon+0x4f6/0x770
unmap_folio+0x196/0x1f0
split_huge_page_to_list_to_order+0x9f6/0x1560
deferred_split_scan+0xac5/0x12a0
shrinker_debugfs_scan_write+0x376/0x470
full_proxy_write+0x15c/0x220
vfs_write+0x2fc/0xcb0
ksys_write+0x146/0x250
do_syscall_64+0x6a/0x120
entry_SYSCALL_64_after_hwframe+0x76/0x7e

The bug is found by syzkaller on an internal kernel, then confirmed on
upstream.

Link: https://lkml.kernel.org/r/20250421113536.3682201-1-gavinguo@igalia.com
Link: https://lore.kernel.org/all/20250414072737.1698513-1-gavinguo@igalia.com/
Link: https://lore.kernel.org/all/20250418085802.2973519-1-gavinguo@igalia.com/
Fixes: 84c3fc4 ("mm: thp: check pmd migration entry in common path")
Signed-off-by: Gavin Guo <gavinguo@igalia.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Gavin Shan <gshan@redhat.com>
Cc: Florent Revest <revest@google.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/huge_memory.c

@@ -3075,6 +3075,8 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
 			   pmd_t *pmd, bool freeze, struct folio *folio)
 {
+	bool pmd_migration = is_pmd_migration_entry(*pmd);
+
 	VM_WARN_ON_ONCE(folio && !folio_test_pmd_mappable(folio));
 	VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));
 	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
@@ -3085,9 +3087,12 @@ void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
 	 * require a folio to check the PMD against. Otherwise, there
 	 * is a risk of replacing the wrong folio.
 	 */
-	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
-	    is_pmd_migration_entry(*pmd)) {
-		if (folio && folio != pmd_folio(*pmd))
+	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || pmd_migration) {
+		/*
+		 * Do not apply pmd_folio() to a migration entry; and folio lock
+		 * guarantees that it must be of the wrong folio anyway.
+		 */
+		if (folio && (pmd_migration || folio != pmd_folio(*pmd)))
 			return;
 		__split_huge_pmd_locked(vma, pmd, address, freeze);
 	}