On Tue, May 07, 2013 at 02:20:03PM -0700, Dave Hansen wrote: > > From: Dave Hansen <[email protected]> > > This was a suggestion from Mel: > > http://lkml.kernel.org/r/[email protected] > > Any pages we collect on 'batch_for_mapping_removal' will have > their lock_page() held during the duration of their stay on the > list. If some other user is trying to get at them during this > time, they might end up having to wait for a while, especially if > we go off and do pageout() on some other page. > > This ensures that we drain the batch if we are about to perform a > writeout. > > I added some statistics to the __remove_mapping_batch() code to > track how large the lists are that we pass in to it. With this > patch, the average list length drops about 10% (from about 4.1 to > 3.8). The workload here was a make -j4 kernel compile on a VM > with 200MB of RAM. > > I've still got the statistics patch around if anyone is > interested. > > Signed-off-by: Dave Hansen <[email protected]> > --- > > linux.git-davehans/kernel/sched/fair.c | 2 ++ > linux.git-davehans/mm/vmscan.c | 10 ++++++++++ > 2 files changed, 12 insertions(+) > > diff -puN kernel/sched/fair.c~drain-batch-list-during-long-operations > kernel/sched/fair.c > --- linux.git/kernel/sched/fair.c~drain-batch-list-during-long-operations > 2013-05-07 13:48:15.267113941 -0700 > +++ linux.git-davehans/kernel/sched/fair.c 2013-05-07 13:48:15.275114295 > -0700 > @@ -5211,6 +5211,8 @@ more_balance: > if (sd->balance_interval < sd->max_interval) > sd->balance_interval *= 2; > } > + //if (printk_ratelimit()) > + // printk("sd->balance_interval: %d\n", sd->balance_interval); > > goto out; >
heh > diff -puN mm/vmscan.c~drain-batch-list-during-long-operations mm/vmscan.c > --- linux.git/mm/vmscan.c~drain-batch-list-during-long-operations > 2013-05-07 13:48:15.268113985 -0700 > +++ linux.git-davehans/mm/vmscan.c 2013-05-07 13:48:15.272114163 -0700 > @@ -936,6 +936,16 @@ static unsigned long shrink_page_list(st > if (!sc->may_writepage) > goto keep_locked; > > + /* > + * We hold a bunch of page locks on the batch. > + * pageout() can take a while, so drain the > + * batch before we perform pageout. > + */ > + nr_reclaimed += > + > __remove_mapping_batch(&batch_for_mapping_removal, > + &ret_pages, > + &free_pages); > + There is also a wait_on_page_writeback() above that would affect memcg's but this alone alleviates a lot of my concerns about lock hold times. -- Mel Gorman SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
