Re: [PATCH] mm/userfaultfd: fix missing PTE unmap for non-migration entries

[David Hildenbrand <david@redhat.com>]

On 01.08.25 16:06, David Hildenbrand wrote:
> On 01.08.25 15:26, Sasha Levin wrote:
>> On Thu, Jul 31, 2025 at 02:56:25PM +0200, David Hildenbrand wrote:
>>> On 31.07.25 14:37, Sasha Levin wrote:
>>>> On Tue, Jul 08, 2025 at 05:42:16PM +0200, David Hildenbrand wrote:
>>>>> On 08.07.25 17:33, Sasha Levin wrote:
>>>>>> On Tue, Jul 08, 2025 at 05:10:44PM +0200, David Hildenbrand wrote:
>>>>>>> On 01.07.25 02:57, Andrew Morton wrote:
>>>>>>>> On Sun, 29 Jun 2025 23:19:58 -0400 Sasha Levin <sashal@kernel.org> wrote:
>>>>>>>>
>>>>>>>>> When handling non-swap entries in move_pages_pte(), the error handling
>>>>>>>>> for entries that are NOT migration entries fails to unmap the page table
>>>>>>>>> entries before jumping to the error handling label.
>>>>>>>>>
>>>>>>>>> This results in a kmap/kunmap imbalance which on CONFIG_HIGHPTE systems
>>>>>>>>> triggers a WARNING in kunmap_local_indexed() because the kmap stack is
>>>>>>>>> corrupted.
>>>>>>>>>
>>>>>>>>> Example call trace on ARM32 (CONFIG_HIGHPTE enabled):
>>>>>>>>>     WARNING: CPU: 1 PID: 633 at mm/highmem.c:622 kunmap_local_indexed+0x178/0x17c
>>>>>>>>>     Call trace:
>>>>>>>>>       kunmap_local_indexed from move_pages+0x964/0x19f4
>>>>>>>>>       move_pages from userfaultfd_ioctl+0x129c/0x2144
>>>>>>>>>       userfaultfd_ioctl from sys_ioctl+0x558/0xd24
>>>>>>>>>
>>>>>>>>> The issue was introduced with the UFFDIO_MOVE feature but became more
>>>>>>>>> frequent with the addition of guard pages (commit 7c53dfbdb024 ("mm: add
>>>>>>>>> PTE_MARKER_GUARD PTE marker")) which made the non-migration entry code
>>>>>>>>> path more commonly executed during userfaultfd operations.
>>>>>>>>>
>>>>>>>>> Fix this by ensuring PTEs are properly unmapped in all non-swap entry
>>>>>>>>> paths before jumping to the error handling label, not just for migration
>>>>>>>>> entries.
>>>>>>>>
>>>>>>>> I don't get it.
>>>>>>>>
>>>>>>>>> --- a/mm/userfaultfd.c
>>>>>>>>> +++ b/mm/userfaultfd.c
>>>>>>>>> @@ -1384,14 +1384,15 @@ static int move_pages_pte(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
>>>>>>>>>    		entry = pte_to_swp_entry(orig_src_pte);
>>>>>>>>>    		if (non_swap_entry(entry)) {
>>>>>>>>> +			pte_unmap(src_pte);
>>>>>>>>> +			pte_unmap(dst_pte);
>>>>>>>>> +			src_pte = dst_pte = NULL;
>>>>>>>>>    			if (is_migration_entry(entry)) {
>>>>>>>>> -				pte_unmap(src_pte);
>>>>>>>>> -				pte_unmap(dst_pte);
>>>>>>>>> -				src_pte = dst_pte = NULL;
>>>>>>>>>    				migration_entry_wait(mm, src_pmd, src_addr);
>>>>>>>>>    				err = -EAGAIN;
>>>>>>>>> -			} else
>>>>>>>>> +			} else {
>>>>>>>>>    				err = -EFAULT;
>>>>>>>>> +			}
>>>>>>>>>    			goto out;
>>>>>>>>
>>>>>>>> where we have
>>>>>>>>
>>>>>>>> out:
>>>>>>>> 	...
>>>>>>>> 	if (dst_pte)
>>>>>>>> 		pte_unmap(dst_pte);
>>>>>>>> 	if (src_pte)
>>>>>>>> 		pte_unmap(src_pte);
>>>>>>>
>>>>>>> AI slop?
>>>>>>
>>>>>> Nah, this one is sadly all me :(
>>>>>
>>>>> Haha, sorry :P
>>>>
>>>> So as I was getting nowhere with this, I asked AI to help me :)
>>>>
>>>> If you're not interested in reading LLM generated code, feel free to
>>>> stop reading now...
>>>>
>>>> After it went over the logs, and a few prompts to point it the right
>>>> way, it ended up generating a patch (below) that made sense, and fixed
>>>> the warning that LKFT was being able to trigger.
>>>>
>>>> If anyone who's more familiar with the code than me (and the AI) agrees
>>>> with the patch and ways to throw their Reviewed-by, I'll send out the
>>>> patch.
>>>
>>> Seems to check out for me. In particular, out pte_unmap() everywhere
>>> else in that function (and mremap.c:move_ptes) are ordered properly.
>>>
>>> Even if it would not fix the issue, it would be a cleanup :)
>>>
>>> Acked-by: David Hildenbrand <david@redhat.com>
>>
>> David, I ended up LLM generating a .cocci script to detect this type of
>> issues, and it ended up detecting a similar issue in
>> arch/loongarch/mm/init.c.
> 
> Does loongarch have these kmap_local restrictions?

loongarch doesn't use HIGHMEM, so it probably doesn't matter. Could be 
considered a cleanup, though.

-- 
Cheers,

David / dhildenb