Re: [PATCH 1/1] mm/madvise: initialize prev pointer in madvise_walk_vmas

[David Hildenbrand <david@redhat.com>]

On 17.06.25 10:50, Lorenzo Stoakes wrote:
> On Tue, Jun 17, 2025 at 10:38:07AM +0200, David Hildenbrand wrote:
>>>>>> 	vma = vma_modify_flags_name(&vmi, *prev, ...
>>>>>>
>>>>>> We should use Fixes: then.
>>>>>
>>>>> So no we shouldn't...
>>>>>
>>>>>>
>>>>>>
>>>>>> Acked-by: David Hildenbrand <david@redhat.com>
>>>>>
>>>>> Sure? :)
>>>>
>>>> Unless I am missing something important, yes :)
>>>
>>> This solution isn't correct as prev == NULL when prev != NULL is wholly
>>> incorrect.
>>
>> I am not able to understand what you mean :)
>>
>> I assume you mean, that we reach a point down in the callchain, where "prev"
>> is supposed to be set to something proper, but it would be "NULL".
> 
> I mean if you tell merge code 'hey the previous VMA is NULL' (same thing as
> saying 'hey this is the first VMA in the address space) and it isn't, bad things
> will happen (TM).
> 
>>
>> That would indeed require a different fix.
> 
> Yes this patch is wrong, sorry.
> 
>>
>> I wonder why we didn't trigger this case so far?
> 
> It's because it only happens since Barry's per-VMA lock logic...
> 
> 	if (madv_behavior && madv_behavior->lock_mode == MADVISE_VMA_READ_LOCK) {
> 		vma = try_vma_read_lock(mm, madv_behavior, start, end);
> 		if (vma) {
> 			error = visit(vma, &prev, start, end, arg);
> 			vma_end_read(vma);
> 			return error;
> 		}
> 	}
> 
> Otherwise, we look up the find_vma_prev():
> 
> 	vma = find_vma_prev(mm, start, &prev);
> 
> In madvise_dontneed_free() we always set *prev = vma _first_.
> 
> Let me suggest the better fix to Lance higher in thread so he sees :)
> 
> Not sure if a fixes is valid here given this isn't mainline yet, more so this
> should be squashed with barry's series?

If it's not in mm-stable yet, it can still be squashed, yes.

-- 
Cheers,

David / dhildenb