Re: [PATCH] mm/madvise: allow MADV_DONTNEED to free memory that is MLOCK_ONFAULT

[Jason Baron]

On 06/11/2018 11:03 AM, Michal Hocko wrote:
> On Mon 11-06-18 10:51:44, Jason Baron wrote:
>> On 06/11/2018 03:20 AM, Michal Hocko wrote:
>>> [CCing linux-api - please make sure to CC this mailing list anytime you
>>>  are touching user visible apis]
>>>
>>> On Fri 08-06-18 14:56:52, Jason Baron wrote:
>>>> In order to free memory that is marked MLOCK_ONFAULT, the memory region
>>>> needs to be first unlocked, before calling MADV_DONTNEED. And if the region
>>>> is to be reused as MLOCK_ONFAULT, we require another call to mlock2() with
>>>> the MLOCK_ONFAULT flag.
>>>>
>>>> Let's simplify freeing memory that is set MLOCK_ONFAULT, by allowing
>>>> MADV_DONTNEED to work directly for memory that is set MLOCK_ONFAULT.
>>>
>>> I do not understand the point here. How is MLOCK_ONFAULT any different
>>> from the regular mlock here? If you want to free mlocked memory then
>>> fine but the behavior should be consistent. MLOCK_ONFAULT is just a way
>>> to say that we do not want to pre-populate the mlocked area and do that
>>> lazily on the page fault time. madvise should make any difference here.
>>>
>>
>> The difference for me is after the page has been freed, MLOCK_ONFAULT
>> will re-populate the range if its accessed again. Whereas with regular
>> mlock I don't think it will because its normally done at mlock() or
>> mmap() time.
> 
> The vma would still be locked so we would effectively turn it into
> ONFAULT IIRC.
>

Indeed. I just tried allowing MADV_DONTNEED against regular mlock() and
in my brief testing it seemed to work as expected against both anonymous
and file back pages. I am certainly not against allowing it for regular
mlock() as well, if you think that makes it more consistent.


>> In any case, the state of a region being locked with
>> regular mlock and pages not present does not currently exist, whereas it
>> does for MLOCK_ONFAULT, so it seems more natural to do it only for
>> MLOCK_ONFAULT. Finally, the use-case we had for this, didn't need
>> regular mlock().
> 
> So can we start discussing whether we want to allow MADV_DONTNEED on
> mlocked areas and what downsides it might have? Sure it would turn the
> strong mlock guarantee to have the whole vma resident but is this
> acceptable for something that is an explicit request from the owner of
> the memory?
> 

If its being explicity requested by the owner it makes sense to me. I
guess there could be a concern about this breaking some userspace that
relied on MADV_DONTNEED not freeing locked memory?

Thanks,

-Jason