Re: [PATCH v1] mm/mremap: fix WARN with uffd that has remap events disabled

[Peter Xu <peterx@redhat.com>]

On Mon, Aug 18, 2025 at 07:53:58PM +0200, David Hildenbrand wrote:
> Registering userfaultd on a VMA that spans at least one PMD and then
> mremap()'ing that VMA can trigger a WARN when recovering from a failed
> page table move due to a page table allocation error.
> 
> The code ends up doing the right thing (recurse, avoiding moving actual
> page tables), but triggering that WARN is unpleasant:
> 
> WARNING: CPU: 2 PID: 6133 at mm/mremap.c:357 move_normal_pmd mm/mremap.c:357 [inline]
> WARNING: CPU: 2 PID: 6133 at mm/mremap.c:357 move_pgt_entry mm/mremap.c:595 [inline]
> WARNING: CPU: 2 PID: 6133 at mm/mremap.c:357 move_page_tables+0x3832/0x44a0 mm/mremap.c:852
> Modules linked in:
> CPU: 2 UID: 0 PID: 6133 Comm: syz.0.19 Not tainted 6.17.0-rc1-syzkaller-00004-g53e760d89498 #0 PREEMPT(full)
> Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2~bpo12+1 04/01/2014
> RIP: 0010:move_normal_pmd mm/mremap.c:357 [inline]
> RIP: 0010:move_pgt_entry mm/mremap.c:595 [inline]
> RIP: 0010:move_page_tables+0x3832/0x44a0 mm/mremap.c:852
> Code: ...
> RSP: 0018:ffffc900037a76d8 EFLAGS: 00010293
> RAX: 0000000000000000 RBX: 0000000032930007 RCX: ffffffff820c6645
> RDX: ffff88802e56a440 RSI: ffffffff820c7201 RDI: 0000000000000007
> RBP: ffff888037728fc0 R08: 0000000000000007 R09: 0000000000000000
> R10: 0000000032930007 R11: 0000000000000000 R12: 0000000000000000
> R13: ffffc900037a79a8 R14: 0000000000000001 R15: dffffc0000000000
> FS:  000055556316a500(0000) GS:ffff8880d68bc000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 0000001b30863fff CR3: 0000000050171000 CR4: 0000000000352ef0
> Call Trace:
>  <TASK>
>  copy_vma_and_data+0x468/0x790 mm/mremap.c:1215
>  move_vma+0x548/0x1780 mm/mremap.c:1282
>  mremap_to+0x1b7/0x450 mm/mremap.c:1406
>  do_mremap+0xfad/0x1f80 mm/mremap.c:1921
>  __do_sys_mremap+0x119/0x170 mm/mremap.c:1977
>  do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
>  do_syscall_64+0xcd/0x4c0 arch/x86/entry/syscall_64.c:94
>  entry_SYSCALL_64_after_hwframe+0x77/0x7f
> RIP: 0033:0x7f00d0b8ebe9
> Code: ...
> RSP: 002b:00007ffe5ea5ee98 EFLAGS: 00000246 ORIG_RAX: 0000000000000019
> RAX: ffffffffffffffda RBX: 00007f00d0db5fa0 RCX: 00007f00d0b8ebe9
> RDX: 0000000000400000 RSI: 0000000000c00000 RDI: 0000200000000000
> RBP: 00007ffe5ea5eef0 R08: 0000200000c00000 R09: 0000000000000000
> R10: 0000000000000003 R11: 0000000000000246 R12: 0000000000000002
> R13: 00007f00d0db5fa0 R14: 00007f00d0db5fa0 R15: 0000000000000005
>  </TASK>
> 
> The underlying issue is that we recurse during the original page table
> move, but not during the recovery move.
> 
> Fix it by checking for both VMAs and performing the check before the
> pmd_none() sanity check.
> 
> Add a new helper where we perform+document that check for the PMD and
> PUD level.
> 
> Thanks to Harry for bisecting.
> 
> Reported-by: syzbot+4d9a13f0797c46a29e42@syzkaller.appspotmail.com
> Closes: https://lkml.kernel.org/r/689bb893.050a0220.7f033.013a.GAE@google.com
> Fixes: 0cef0bb836e ("mm: clear uffd-wp PTE/PMD state on mremap()")
> Cc: <stable@kernel.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
> Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Cc: Jann Horn <jannh@google.com>
> Cc: Pedro Falcato <pfalcato@suse.de>
> Cc: Harry Yoo <harry.yoo@oracle.com>
> Signed-off-by: David Hildenbrand <david@redhat.com>

Acked-by: Peter Xu <peterx@redhat.com>

-- 
Peter Xu