[PATCH v2] mm: support anonymous stable page

[Minchan Kim]

On Fri, Nov 18, 2016 at 01:26:59PM -0800, Hugh Dickins wrote:
> On Fri, 18 Nov 2016, Minchan Kim wrote:
> > On Thu, Nov 17, 2016 at 08:35:10PM -0800, Hugh Dickins wrote:
> > > 
> > > Maybe add SWP_STABLE_WRITES in include/linux/swap.h, and set that
> > > in swap_info->flags according to bdi_cap_stable_pages_required(),
> > > leaving mapping->host itself NULL as before?
> > 
> > The problem with the approach is that we need to get swap_info_struct
> > in reuse_swap_page so maybe, every caller should pass swp_entry_t
> > into reuse_swap_page. It would be no problem if swap slot is really
> > referenced the page(IOW, pte is real swp_entry_t) but some cases
> > where swap slot is already empty but the page remains in only
> > swap cache, we cannot pass swp_entry_t which means that we cannot
> > get swap_info_struct.
> 
> I don't see the problem: if the page is PageSwapCache (and page
> lock is held), then the swp_entry_t is there in page->private:
> see page_swapcount(), which reuse_swap_page() just called.

Right you are!!

> 
> > 
> > So, if I didn't miss, another option I can imagine is to move
> > SWP_STABLE_WRITES to address_space->flags as AS_STABLE_WRITES.
> > With that, we can always get the information without passing
> > swp_entry_t. Is there any better idea?
> 
> I think what you suggest below would work fine (and be quicker
> than looking up the swap_info again): but is horribly misleading
> for anyone else interested in stable writes, for whom the info is
> kept in the bdi, and not in this new mapping flag.
> 
> So I'd much prefer that you keep the swap special case inside the
> world of swap, with a SWP_STABLE_WRITES flag.  Maybe unfold
> page_swapcount() inside reuse_swap_page(), so that it only
> needs a single lookup (or perhaps I'm optimizing prematurely).
> 

I toally agree.
Here is new one.