Re: [patch 07/11] mm: vmscan: convert unevictable page rescue scanner to per-memcg LRU lists

[Johannes Weiner <jweiner@redhat.com>]

On Wed, Sep 21, 2011 at 02:33:56PM +0200, Michal Hocko wrote:
> On Mon 12-09-11 12:57:24, Johannes Weiner wrote:
> > The global per-zone LRU lists are about to go away on memcg-enabled
> > kernels, the unevictable page rescue scanner must be able to find its
> > pages on the per-memcg LRU lists.
> > 
> > Signed-off-by: Johannes Weiner <jweiner@redhat.com>
> 
> The patch is correct but I guess the original implementation of
> scan_zone_unevictable_pages is buggy (see bellow). This should be
> addressed separatelly, though.
> 
> Reviewed-by: Michal Hocko <mhocko@suse.cz>

Thanks for your effort, Michal, I really appreciate it.

> > @@ -3490,32 +3501,40 @@ void scan_mapping_unevictable_pages(struct address_space *mapping)
> >  #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
> >  static void scan_zone_unevictable_pages(struct zone *zone)
> >  {
> > -	struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
> > -	unsigned long scan;
> > -	unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
> > -
> > -	while (nr_to_scan > 0) {
> > -		unsigned long batch_size = min(nr_to_scan,
> > -						SCAN_UNEVICTABLE_BATCH_SIZE);
> > -
> > -		spin_lock_irq(&zone->lru_lock);
> > -		for (scan = 0;  scan < batch_size; scan++) {
> > -			struct page *page = lru_to_page(l_unevictable);
> > +	struct mem_cgroup *mem;
> >  
> > -			if (!trylock_page(page))
> > -				continue;
> > +	mem = mem_cgroup_iter(NULL, NULL, NULL);
> > +	do {
> > +		struct mem_cgroup_zone mz = {
> > +			.mem_cgroup = mem,
> > +			.zone = zone,
> > +		};
> > +		unsigned long nr_to_scan;
> >  
> > -			prefetchw_prev_lru_page(page, l_unevictable, flags);
> > +		nr_to_scan = zone_nr_lru_pages(&mz, LRU_UNEVICTABLE);
> > +		while (nr_to_scan > 0) {
> > +			unsigned long batch_size;
> > +			unsigned long scan;
> >  
> > -			if (likely(PageLRU(page) && PageUnevictable(page)))
> > -				check_move_unevictable_page(page, zone);
> > +			batch_size = min(nr_to_scan,
> > +					 SCAN_UNEVICTABLE_BATCH_SIZE);
> > +			spin_lock_irq(&zone->lru_lock);
> > +			for (scan = 0; scan < batch_size; scan++) {
> > +				struct page *page;
> >  
> > -			unlock_page(page);
> > +				page = lru_tailpage(&mz, LRU_UNEVICTABLE);
> > +				if (!trylock_page(page))
> > +					continue;
> 
> We are not moving to the next page so we will try it again in the next
> round while we already increased the scan count. In the end we will
> missed some pages.

I guess this is about latency.  This code is only executed when the
user requests so by writing to a proc-file, check the comment above
scan_all_zones_unevictable_pages.  I think at one point Lee wanted to
move anon pages to the unevictable LRU when no swap is configured, but
we have separate anon LRUs now that are not scanned without swap, and
I think except for bugs there is no actual need to move these pages by
hand, let alone reliably every single page.