Re: [RFC] Tracking mlocked pages and moving them off the LRU

["Martin J. Bligh" <mbligh@mbligh.org>]

Christoph Hellwig wrote:
> On Fri, Feb 02, 2007 at 10:20:12PM -0800, Christoph Lameter wrote:
>> This is a new variation on the earlier RFC for tracking mlocked pages.
>> We now mark a mlocked page with a bit in the page flags and remove
>> them from the LRU. Pages get moved back when no vma that references
>> the page has VM_LOCKED set anymore.
>>
>> This means that vmscan no longer uselessly cycles over large amounts
>> of mlocked memory should someone attempt to mlock large amounts of
>> memory (may even result in a livelock on large systems).
>>
>> Synchronization is build around state changes of the PageMlocked bit.
>> The NR_MLOCK counter is incremented and decremented based on
>> state transitions of PageMlocked. So the count is accurate.
>>
>> There is still some unfinished business:
>>
>> 1. We use the 21st page flag and we only have 20 on 32 bit NUMA platforms.
>>
>> 2. Since mlocked pages are now off the LRU page migration will no longer
>>    move them.
>>
>> 3. Use NR_MLOCK to tune various VM behaviors so that the VM does not 
>>    longer fall due to too many mlocked pages in certain areas.
> 
> This patch seems to not handle the cases where more than one process mlocks
> a page and you really need a pincount in the page to not release it before
> all processes have munlock it or died.  I did a similar patch a while
> ago and tried to handle it by overloading the lru lists pointers with
> a pincount, but at some point I gave up because I couldn't get that part
> right.

Doesn't matter - you can just do it lazily. If you find a page that is
locked, move it to the locked list. when unlocking a page you *always*
move it back to the normal list. If someone else is still locking it,
we'll move it back to the lock list on next reclaim pass.

I have a half-finished patch from 6 months ago that does this, but never
found time to complete it ;-(

M.