Re: [RFC PATCH v1 12/13] mm: split up release_pages into non-sentinel and sentinel passes

[Daniel Jordan]

On 02/02/2018 12:00 PM, Laurent Dufour wrote:
> On 02/02/2018 15:40, Laurent Dufour wrote:
> > On 01/02/2018 00:04, daniel.m.jordan@xxxxxxxxxx wrote:
> > > A common case in release_pages is for the 'pages' list to be in roughly
> > > the same order as they are in their LRU.  With LRU batch locking, when a
> > > sentinel page is removed, an adjacent non-sentinel page must be promoted
> > > to a sentinel page to follow the locking scheme.  So we can get behavior
> > > where nearly every page in the 'pages' array is treated as a sentinel
> > > page, hurting the scalability of this approach.
> > >
> > > To address this, split up release_pages into non-sentinel and sentinel
> > > passes so that the non-sentinel pages can be locked with an LRU batch
> > > lock before the sentinel pages are removed.
> > >
> > > For the prototype, just use a bitmap and a temporary outer loop to
> > > implement this.
> > >
> > > Performance numbers from a single microbenchmark at this point in the
> > > series are included in the next patch.
> > >
> > > Signed-off-by: Daniel Jordan <daniel.m.jordan@xxxxxxxxxx>
> > > ---
> > >   mm/swap.c | 20 +++++++++++++++++++-
> > >   1 file changed, 19 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/mm/swap.c b/mm/swap.c
> > > index fae766e035a4..a302224293ad 100644
> > > --- a/mm/swap.c
> > > +++ b/mm/swap.c
> > > @@ -731,6 +731,7 @@ void lru_add_drain_all(void)
> > >   	put_online_cpus();
> > >   }
> > >
> > > +#define LRU_BITMAP_SIZE	512
> > >   /**
> > >    * release_pages - batched put_page()
> > >    * @pages: array of pages to release
> > > @@ -742,16 +743,32 @@ void lru_add_drain_all(void)
> > >    */
> > >   void release_pages(struct page **pages, int nr)
> > >   {
> > > -	int i;
> > > +	int h, i;
> > >   	LIST_HEAD(pages_to_free);
> > >   	struct pglist_data *locked_pgdat = NULL;
> > >   	spinlock_t *locked_lru_batch = NULL;
> > >   	struct lruvec *lruvec;
> > >   	unsigned long uninitialized_var(flags);
> > > +	DECLARE_BITMAP(lru_bitmap, LRU_BITMAP_SIZE);
> > > +
> > > +	VM_BUG_ON(nr > LRU_BITMAP_SIZE);
> >
> > While running your series rebased on v4.15-mmotm-2018-01-31-16-51, I'm
> > hitting this VM_BUG sometimes on a ppc64 system where page size is set to 64K.
>
> I can't see any link between nr and LRU_BITMAP_SIZE, caller may pass a
> larger list of pages which is not relative to the LRU list.

You're correct, I used the hard-coded size to quickly prototype, just to 
see how this approach performs.  That's unfortunate that it bit you.
 > To move forward seeing the benefit of this series with the SPF one, I
> declared the bit map based on nr. This is still not a valid option but this
> at least allows to process all the passed pages.

Yes, the bitmap's not for the final version.