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

From: Jason Baron
Date: Fri Jun 15 2018 - 15:29:20 EST

On 06/13/2018 05:13 AM, Michal Hocko wrote:
> On Tue 12-06-18 10:11:33, Jason Baron wrote:
> [...]
>> Ok, I share the concern that there is a chance that userspace is relying
>> on MADV_DONTNEED not free'ing locked memory. In that case, what if we
>> introduce a MADV_DONTNEED_FORCE, which does everything that
>> MADV_DONTNEED currently does but in addition will also free mlock areas.
> What about other types of vmas that are not allowed to MADV_DONTNEED?
> _FORCE suggests that the user of the API know what he is doing so why
> shouldn't we allow unmapping hugetlb pages or special PFNMAPS? Or do we
> want to add MADV_DONTNEED_FORCE_FOR_REAL when somebody comes with
> another usecase?
> I agree with Vlastimil that adding new madvise mode for niche case
> sounds like a bad idea so we should better be sure that a new flag has
> a reasonable semantic. Just allow mlocked pages is more of a tweak than
> a proper semantic. So making it force for real requires to analyze what
> that would mean for other vmas which are excluded now.

If its a new flag, I agree it makes sense to look at hugetlb and
pfnmaps. pfnmaps might be more work, since it may require callbacks to
do driver specific actions...

I was able to do something very close to the original requirement of
free'ing mlock'd pages, via using a tmpfs mmap that is mlock'd. And then
using fallocate(FALLOC_FL_PUNCH_HOLE) to free the locked pages. I think
the tmpfs is sufficient for my needs, I wonder if there is any other
interest in this feature?