Re: [RFC][PATCH 4/4] vmscan: vmscan don't use pcp list

From: Mel Gorman
Date: Wed Dec 02 2009 - 09:25:47 EST

On Wed, Dec 02, 2009 at 04:15:37PM +0900, KOSAKI Motohiro wrote:
> Hi
>
> sorry for the delayed reply. I've got stucked in Larry's serious bug report awhile.
>

No worries. I am slow to respond at the best of times.

> > On Fri, Nov 27, 2009 at 09:23:57AM +0900, KOSAKI Motohiro wrote:
> > >
> > > note: Last year, Andy Whitcroft reported pcp prevent to make contenious
> > > high order page when lumpy reclaim is running.
> >
> > I don't remember the specifics of the discussion but I know that when
> > that patch series was being prototyped, it was because order-0
> > allocations were racing with lumpy reclaimers. A lumpy reclaim might
> > free up an order-9 page say but while it was freeing, an order-0 page
> > would be allocated from the middle. It wasn't the PCP lists as such that
> > were a problem once they were getting drained as part of a high-order
> > allocation attempt. It would be just as bad if the order-0 page was
> > taken from the buddy lists.
>
> Hm, probably I have to update my patch description.
> if we use pavevec_free(), batch size is PAGEVEC_SIZE(=14).
> then, order-9 lumpy reclaim makes 37 times pagevec_free(). it makes lots
> temporary uncontenious memory block and the chance of stealing it from
> order-0 allocator task.
>

Very true. It opens a wide window during with other allocation requests
can race with the lumpy reclaimer and undo their work.

> This patch free all reclaimed pages at once to buddy.
>

Which is good. It reduces the window during which trouble can happen
considerably.

> > > He posted "capture pages freed during direct reclaim for allocation by the reclaimer"
> > > patch series, but Christoph mentioned simple bypass pcp instead.
> > > I made it. I'd hear Christoph and Mel's mention.
> > >
> > > ==========================
> > > Currently vmscan free unused pages by __pagevec_free(). It mean free pages one by one
> > > and use pcp. it makes two suboptimal result.
> > >
> > > - The another task can steal the freed page in pcp easily. it decrease
> > > lumpy reclaim worth.
> > > - To pollute pcp cache, vmscan freed pages might kick out cache hot
> > > pages from pcp.
> > >
> >
> > The latter point is interesting.
>
> Thank you.
>

Another point is that lumpy reclaim releases pages via the PCP means that
a part of the contiguous page is "stuck" in the PCP lists. This is evaded
by doing a drain_all_pages() for high-order allocation requests that are
failing. I suspect that your patch will reduce the number of times the PCP
lists are drained.

> > > This patch make new free_pages_bulk() function and vmscan use it.
> > >
> > > Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx>
> > > ---
> > > include/linux/gfp.h | 2 +
> > > mm/page_alloc.c | 56 +++++++++++++++++++++++++++++++++++++++++++++++++++
> > > mm/vmscan.c | 23 +++++++++++----------
> > > 3 files changed, 70 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > > index f53e9b8..403584d 100644
> > > --- a/include/linux/gfp.h
> > > +++ b/include/linux/gfp.h
> > > @@ -330,6 +330,8 @@ extern void free_hot_page(struct page *page);
> > > #define __free_page(page) __free_pages((page), 0)
> > > #define free_page(addr) free_pages((addr),0)
> > >
> > > +void free_pages_bulk(struct zone *zone, int count, struct list_head *list);
> > > +
> > > void page_alloc_init(void);
> > > void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp);
> > > void drain_all_pages(void);
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 11ae66e..f77f8a8 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -2037,6 +2037,62 @@ void free_pages(unsigned long addr, unsigned int order)
> > >
> > > EXPORT_SYMBOL(free_pages);
> > >
> > > +/*
> > > + * Frees a number of pages from the list
> > > + * Assumes all pages on list are in same zone and order==0.
> > > + * count is the number of pages to free.
> > > + *
> > > + * This is similar to __pagevec_free(), but receive list instead pagevec.
> > > + * and this don't use pcp cache. it is good characteristics for vmscan.
> > > + */
> > > +void free_pages_bulk(struct zone *zone, int count, struct list_head *list)
> > > +{
> > > + unsigned long flags;
> > > + struct page *page;
> > > + struct page *page2;
> > > +
> > > + list_for_each_entry_safe(page, page2, list, lru) {
> > > + int wasMlocked = __TestClearPageMlocked(page);
> > > +
> > > + kmemcheck_free_shadow(page, 0);
> > > +
> > > + if (PageAnon(page))
> > > + page->mapping = NULL;
> > > + if (free_pages_check(page)) {
> > > + /* orphan this page. */
> > > + list_del(&page->lru);
> > > + continue;
> > > + }
> > > + if (!PageHighMem(page)) {
> > > + debug_check_no_locks_freed(page_address(page),
> > > + PAGE_SIZE);
> > > + debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
> > > + }
> > > + arch_free_page(page, 0);
> > > + kernel_map_pages(page, 1, 0);
> > > +
> > > + local_irq_save(flags);
> > > + if (unlikely(wasMlocked))
> > > + free_page_mlock(page);
> > > + local_irq_restore(flags);
> > > + }
> > > +
> > > + spin_lock_irqsave(&zone->lock, flags);
> > > + __count_vm_events(PGFREE, count);
> > > + zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
> > > + zone->pages_scanned = 0;
> > > +
> > > + __mod_zone_page_state(zone, NR_FREE_PAGES, count);
> > > +
> > > + list_for_each_entry_safe(page, page2, list, lru) {
> > > + /* have to delete it as __free_one_page list manipulates */
> > > + list_del(&page->lru);
> > > + trace_mm_page_free_direct(page, 0);
> > > + __free_one_page(page, zone, 0, page_private(page));
> > > + }
> > > + spin_unlock_irqrestore(&zone->lock, flags);
> > > +}
> >
> > It would be preferable that the bulk free code would use as much of the
> > existing free logic in the page allocator as possible. This is making a
> > lot of checks that are done elsewhere. As this is an RFC, it's not
> > critical but worth bearing in mind.
>
> Sure. I have to merge common block. thanks.
>
>
> > > +
> > > /**
> > > * alloc_pages_exact - allocate an exact number physically-contiguous pages.
> > > * @size: the number of bytes to allocate
> > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > index 56faefb..00156f2 100644
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -598,18 +598,17 @@ redo:
> > > * shrink_page_list() returns the number of reclaimed pages
> > > */
> > > static unsigned long shrink_page_list(struct list_head *page_list,
> > > + struct list_head *freed_pages_list,
> > > struct scan_control *sc,
> >
> > Should the freed_pages_list be part of scan_control?
>
> OK.
>
> >
> > > enum pageout_io sync_writeback)
> > > {
> > > LIST_HEAD(ret_pages);
> > > - struct pagevec freed_pvec;
> > > int pgactivate = 0;
> > > unsigned long nr_reclaimed = 0;
> > > unsigned long vm_flags;
> > >
> > > cond_resched();
> > >
> > > - pagevec_init(&freed_pvec, 1);
> > > while (!list_empty(page_list)) {
> > > struct address_space *mapping;
> > > struct page *page;
> > > @@ -785,10 +784,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
> > > __clear_page_locked(page);
> > > free_it:
> > > nr_reclaimed++;
> > > - if (!pagevec_add(&freed_pvec, page)) {
> > > - __pagevec_free(&freed_pvec);
> > > - pagevec_reinit(&freed_pvec);
> > > - }
> > > + list_add(&page->lru, freed_pages_list);
> > > continue;
> > >
> > > cull_mlocked:
> > > @@ -812,8 +808,6 @@ keep:
> > > VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
> > > }
> > > list_splice(&ret_pages, page_list);
> > > - if (pagevec_count(&freed_pvec))
> > > - __pagevec_free(&freed_pvec);
> > > count_vm_events(PGACTIVATE, pgactivate);
> > > return nr_reclaimed;
> > > }
> > > @@ -1100,6 +1094,7 @@ static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
> > > int priority, int file)
> > > {
> > > LIST_HEAD(page_list);
> > > + LIST_HEAD(freed_pages_list);
> > > struct pagevec pvec;
> > > unsigned long nr_scanned;
> > > unsigned long nr_reclaimed = 0;
> > > @@ -1174,7 +1169,8 @@ static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
> > >
> > > spin_unlock_irq(&zone->lru_lock);
> > >
> > > - nr_reclaimed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
> > > + nr_reclaimed = shrink_page_list(&page_list, &freed_pages_list, sc,
> > > + PAGEOUT_IO_ASYNC);
> > >
> > > /*
> > > * If we are direct reclaiming for contiguous pages and we do
> > > @@ -1192,10 +1188,15 @@ static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
> > > nr_active = clear_active_flags(&page_list, count);
> > > count_vm_events(PGDEACTIVATE, nr_active);
> > >
> > > - nr_reclaimed += shrink_page_list(&page_list, sc,
> > > - PAGEOUT_IO_SYNC);
> > > + nr_reclaimed += shrink_page_list(&page_list, &freed_pages_list,
> > > + sc, PAGEOUT_IO_SYNC);
> > > }
> > >
> > > + /*
> > > + * Free unused pages.
> > > + */
> > > + free_pages_bulk(zone, nr_reclaimed, &freed_pages_list);
> > > +
> > > local_irq_disable();
> > > if (current_is_kswapd())
> > > __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
> >
> > This patch does not stand-alone so it's not easy to test. I'll think about
> > the idea more although I do see how it might help slightly in the same way
> > capture-reclaim did by closing the race window with other allocators.
> >
> > I'm curious, how did you evaluate this and what problem did you
> > encounter that this might help?
>
> Honestly I didn't it yet. I only tested changing locking scheme didn't cause
> reclaim throughput under light VM pressure. Probably I have to contact
> Andy and test his original problem workload.
>
> btw, if you have good high order allocation workload, can you please tell me it?
>

For the most part, he was using the same tests as I was using for the
anti-fragmentation patches - high-order allocation requests under a heavy
compile-load. The expectation was that capture-based reclaim would increase
success rates and reduce latencies.

--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/