Re: [PATCH 3/5] mm/migrate.c: fix potential indeterminate pte entry in migrate_vma_insert_page()

[Miaohe Lin]

On 2021/3/23 19:28, Miaohe Lin wrote:
> On 2021/3/23 19:07, Alistair Popple wrote:
>> On Tuesday, 23 March 2021 9:26:43 PM AEDT David Hildenbrand wrote:
>>> On 20.03.21 10:36, Miaohe Lin wrote:
>>>> If the zone device page does not belong to un-addressable device memory,
>>>> the variable entry will be uninitialized and lead to indeterminate pte
>>>> entry ultimately. Fix this unexpectant case and warn about it.
>>>
>>> s/unexpectant/unexpected/
>>>
>>>>
>>>> Fixes: df6ad69838fc ("mm/device-public-memory: device memory cache 
>> coherent with CPU")
>>>> Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>>>> ---
>>>>   mm/migrate.c | 7 +++++++
>>>>   1 file changed, 7 insertions(+)
>>>>
>>>> diff --git a/mm/migrate.c b/mm/migrate.c
>>>> index 20a3bf75270a..271081b014cb 100644
>>>> --- a/mm/migrate.c
>>>> +++ b/mm/migrate.c
>>>> @@ -2972,6 +2972,13 @@ static void migrate_vma_insert_page(struct 
>> migrate_vma *migrate,
>>>>   
>>>>   			swp_entry = make_device_private_entry(page, vma->vm_flags & 
>> VM_WRITE);
>>>>   			entry = swp_entry_to_pte(swp_entry);
>>>> +		} else {
>>>> +			/*
>>>> +			 * For now we only support migrating to un-addressable
>>>> +			 * device memory.
>>>> +			 */
>>>> +			WARN_ON(1);
>>>> +			goto abort;
>>>
>>> Fix it by crashing the kernel with panic_on_warn? :)
>>>
>>> If this case can actual happen, than no WARN_ON() - rather a 
>>> pr_warn_once(). If this case cannot happen, why do we even care (it's 
>>> not a fix then)?
>>
>> There is also already a check for this case in migrate_vma_pages(). The 
>> problem is it happens after the call to migrate_vma_insert_page(). I wonder if 
>> instead it would be better just to move the existing check to before that 
>> call?
>>
> 

While look at this more closely, move the existing check to before migrate_vma_insert_page() could
potentially change the current mmu_notifier semantics against anonymous zero page. :(
So check zone device page inside migrate_vma_insert_page() would be more acceptable.
Many thanks.

> Yes, sounds good! Many thanks for your advice! :)
> 
>>>
>> .
>>
>