Re: [RFC 06/11] mm: remove SWAP_MLOCK in ttu

[Minchan Kim]

On Tue, Mar 07, 2017 at 06:24:37PM +0300, Kirill A. Shutemov wrote:
> On Mon, Mar 06, 2017 at 11:15:08AM +0900, Minchan Kim wrote:
> > Hi Anshuman,
> > 
> > On Fri, Mar 03, 2017 at 06:06:38PM +0530, Anshuman Khandual wrote:
> > > On 03/02/2017 12:09 PM, Minchan Kim wrote:
> > > > ttu don't need to return SWAP_MLOCK. Instead, just return SWAP_FAIL
> > > > because it means the page is not-swappable so it should move to
> > > > another LRU list(active or unevictable). putback friends will
> > > > move it to right list depending on the page's LRU flag.
> > > 
> > > Right, if it cannot be swapped out there is not much difference with
> > > SWAP_FAIL once we change the callers who expected to see a SWAP_MLOCK
> > > return instead.
> > > 
> > > > 
> > > > A side effect is shrink_page_list accounts unevictable list movement
> > > > by PGACTIVATE but I don't think it corrupts something severe.
> > > 
> > > Not sure I got that, could you please elaborate on this. We will still
> > > activate the page and put it in an appropriate LRU list if it is marked
> > > mlocked ?
> > 
> > Right. putback_iactive_pages/putback_lru_page has a logic to filter
> > out unevictable pages and move them to unevictable LRU list so it
> > doesn't break LRU change behavior but the concern is until now,
> > we have accounted PGACTIVATE for only evictable LRU list page but
> > by this change, it accounts it to unevictable LRU list as well.
> > However, although I don't think it's big problem in real practice,
> > we can fix it simply with checking PG_mlocked if someone reports.
> 
> I think it's better to do this pro-actively. Let's hide both pgactivate++
> and SetPageActive() under "if (!PageMlocked())".
> SetPageActive() is not free.

I will consider it in next spin.

Thanks!