Re: [PATCH v3 02/16] mm/compaction: support non-lru movable page migration

[Vlastimil Babka]

On 04/04/2016 07:12 AM, Minchan Kim wrote:
> On Fri, Apr 01, 2016 at 11:29:14PM +0200, Vlastimil Babka wrote:
> > Might have been better as a separate migration patch and then a
> > compaction patch. It's prefixed mm/compaction, but most changed are
> > in mm/migrate.c
>
> Indeed. The title is rather misleading but not sure it's a good idea
> to separate compaction and migration part.

Guess it's better to see the new functions together with its user after 
all, OK.

> I will just resend to change the tile from "mm/compaction" to
> "mm/migration".

OK!

> > Also I'm a bit uncomfortable how isolate_movable_page() blindly expects that
> > page->mapping->a_ops->isolate_page exists for PageMovable() pages.
> > What if it's a false positive on a PG_reclaim page? Can we rely on
> > PG_reclaim always (and without races) implying PageLRU() so that we
> > don't even attempt isolate_movable_page()?
>
> For now, we shouldn't have such a false positive because PageMovable
> checks page->_mapcount == PAGE_MOVABLE_MAPCOUNT_VALUE as well as PG_movable
> under PG_lock.
>
> But I read your question about user-mapped drvier pages so we cannot
> use _mapcount anymore so I will find another thing. A option is this.
>
> static inline int PageMovable(struct page *page)
> {
>          int ret = 0;
>          struct address_space *mapping;
>          struct address_space_operations *a_op;
>
>          if (!test_bit(PG_movable, &(page->flags))
>                  goto out;
>
>          mapping = page->mapping;
>          if (!mapping)
>                  goto out;
>
>          a_op = mapping->a_op;
>          if (!aop)
>                  goto out;
>          if (a_op->isolate_page)
>                  ret = 1;
> out:
>          return ret;
>
> }
>
> It works under PG_lock but with this, we need trylock_page to peek
> whether it's movable non-lru or not for scanning pfn.

Hm I hoped that with READ_ONCE() we could do the peek safely without 
trylock_page, if we use it only as a heuristic. But I guess it would 
require at least RCU-level protection of the 
page->mapping->a_op->isolate_page chain.

> For avoiding that, we need another function to peek which just checks
> PG_movable bit instead of all things.
>
>
> /*
>   * If @page_locked is false, we cannot guarantee page->mapping's stability
>   * so just the function checks with PG_movable which could be false positive
>   * so caller should check it again under PG_lock to check a_ops->isolate_page.
>   */
> static inline int PageMovable(struct page *page, bool page_locked)
> {
>          int ret = 0;
>          struct address_space *mapping;
>          struct address_space_operations *a_op;
>
>          if (!test_bit(PG_movable, &(page->flags))
>                  goto out;
>
>          if (!page_locked) {
>                  ret = 1;
>                  goto out;
>          }
>
>          mapping = page->mapping;
>          if (!mapping)
>                  goto out;
>
>          a_op = mapping->a_op;
>          if (!aop)
>                  goto out;
>          if (a_op->isolate_page)
>                  ret = 1;
> out:
>          return ret;
> }

I wouldn't put everything into single function, but create something 
like __PageMovable() just for the unlocked peek. Unlike the 
zone->lru_lock, we don't keep page_lock() across iterations in 
isolate_migratepages_block(), as obviously each page has different lock.
So the page_locked parameter would be always passed as constant, and at 
that point it's better to have separate functions.

So I guess the question is how many false positives from overlap with 
PG_reclaim the scanner will hit if we give up on 
PAGE_MOVABLE_MAPCOUNT_VALUE, as that will increase number of page locks 
just to realize that it's not actual PageMovable() page...

> Thanks for detail review, Vlastimil!
> I will resend new versions after vacation in this week.

You're welcome, great!