Re: [PATCH] mm/memcg: add MAX_CHARGE_BATCH to limit unnecessary charge overhead

[Johannes Weiner <hannes@xxxxxxxxxxx>]

On Sun, Jun 24, 2012 at 06:32:58PM +0800, Wanpeng Li wrote:
> On Sun, Jun 24, 2012 at 12:19:48PM +0200, Johannes Weiner wrote:
> >On Sun, Jun 24, 2012 at 06:08:26PM +0800, Wanpeng Li wrote:
> >> On Sun, Jun 24, 2012 at 11:46:14AM +0200, Johannes Weiner wrote:
> >> >On Sun, Jun 24, 2012 at 10:16:09AM +0800, Wanpeng Li wrote:
> >> >> From: Wanpeng Li <liwp@xxxxxxxxxxxxxxxxxx>
> >> >> 
> >> >> Since exceeded unused cached charges would add pressure to
> >> >> mem_cgroup_do_charge, more overhead would burn cpu cycles when
> >> >> mem_cgroup_do_charge cause page reclaim or even OOM be triggered
> >> >> just for such exceeded unused cached charges. Add MAX_CHARGE_BATCH
> >> >> to limit max cached charges.
> >> >> 
> >> >> Signed-off-by: Wanpeng Li <liwp.linux@xxxxxxxxx>
> >> >> ---
> >> >>  mm/memcontrol.c |   16 ++++++++++++++++
> >> >>  1 file changed, 16 insertions(+)
> >> >> 
> >> >> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> >> >> index 0e092eb..1ff317a 100644
> >> >> --- a/mm/memcontrol.c
> >> >> +++ b/mm/memcontrol.c
> >> >> @@ -1954,6 +1954,14 @@ void mem_cgroup_update_page_stat(struct page *page,
> >> >>   * TODO: maybe necessary to use big numbers in big irons.
> >> >>   */
> >> >>  #define CHARGE_BATCH	32U
> >> >> +
> >> >> +/*
> >> >> + * Max size of charge stock. Since exceeded unused cached charges would
> >> >> + * add pressure to mem_cgroup_do_charge which will cause page reclaim or
> >> >> + * even oom be triggered.
> >> >> + */
> >> >> +#define MAX_CHARGE_BATCH 1024U
> >> >> +
> >> >>  struct memcg_stock_pcp {
> >> >>  	struct mem_cgroup *cached; /* this never be root cgroup */
> >> >>  	unsigned int nr_pages;
> >> >> @@ -2250,6 +2258,7 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
> >> >>  	unsigned int batch = max(CHARGE_BATCH, nr_pages);
> >> >>  	int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
> >> >>  	struct mem_cgroup *memcg = NULL;
> >> >> +	struct memcg_stock_pcp *stock;
> >> >>  	int ret;
> >> >>  
> >> >>  	/*
> >> >> @@ -2320,6 +2329,13 @@ again:
> >> >>  		rcu_read_unlock();
> >> >>  	}
> >> >>  
> >> >> +	stock = &get_cpu_var(memcg_stock);
> >> >> +	if (memcg == stock->cached && stock->nr_pages) {
> >> >> +		if (stock->nr_pages > MAX_CHARGE_BATCH)
> >> >> +			batch = nr_pages;
> >> >> +	}
> >> >> +	put_cpu_var(memcg_stock);
> >> >
> >> >The only way excessive stock can build up is if the charging task gets
> >> >rescheduled, after trying to consume stock a few lines above, to a cpu
> >> >it was running on when it built up stock in the past.
> >> >
> >> >    consume_stock()
> >> >      memcg != stock->cached:
> >> >        return false
> >> >    do_charge()
> >> >    <reschedule>
> >> >    refill_stock()
> >> >      memcg == stock->cached:
> >> >        stock->nr_pages += nr_pages
> >> 
> >> __mem_cgroup_try_charge() {
> >> 	unsigned int batch = max(CHARGE_BATCH, nr_pages);
> >> 	[...]
> >> 	mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
> >> 	[...]
> >> 	if(batch > nr_pages)
> >> 		refill_stock(memcg, batch - nr_pages);
> >> }
> >> 
> >> Consider this scenario, If one task wants to charge nr_pages = 1,
> >> then batch = max(32,1) = 32, this time 31 excess charges 
> >> will be charged in mem_cgroup_do_charge and then add to stock by
> >> refill_stock. Generally there are many tasks in one memory cgroup and 
> >> maybe charges frequency. In this situation, limit will reach soon, 
> >> and cause mem_cgroup_reclaim to call try_to_free_mem_cgroup_pages.
> >
> >But the stock is not a black hole that gets built up for giggles!  The
> >next time the processes want to charge a page on this cpu, they will
> >consume it from the stock.  Not add more pages to it.  Look at where
> >consume_stock() is called.
> 
> if(nr_pages == 1 && consume_stock(memcg))
> 	goto done;
> 
> Only when charge one page will call consume_stock. You can see the codes
> in mem_cgroup_charge_common() which also call __mem_cgroup_try_charge, 
> when both transparent huge and hugetlbfs pages, nr_pages will larger than 1.

In which case, nr_pages will be bigger than CHARGE_BATCH, in which
case batch equals nr_pages, in which case stock won't be refilled:

  unsigned int batch = max(CHARGE_BATCH, nr_pages);
  ...
  if (batch > nr_pages)
    refill_stock(memcg, batch - nr_pages);

We could maybe make this

  if (nr_pages == 1 && batch > nr_pages)
    ...

for clarity, but it won't make a behavioural difference.
--
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/