On Thu, Dec 9, 2010 at 11:01 AM, Andrew Morton
<akpm@linux-foundation.org> wrote:
> On Thu, 9 Dec 2010 10:55:24 +0900 Minchan Kim <minchan.kim@gmail.com> wrote:
>
>> >> > leaves them to direct reclaim.
>> >>
>> >> Hi!
>> >>
>> >> We are experiencing a similar issue, though with a 757 MB Normal zone,
>> >> where kswapd tries to rebalance Normal after an order-3 allocation while
>> >> page cache allocations (order-0) keep splitting it back up again. __It can
>> >> run the whole day like this (SSD storage) without sleeping.
>> >
>> > People at google have told me they've seen the same thing. __A fork is
>> > taking 15 minutes when someone else is doing a dd, because the fork
>> > enters direct-reclaim trying for an order-one page. __It successfully
>> > frees some order-one pages but before it gets back to allocate one, dd
>> > has gone and stolen them, or split them apart.
>> >
>> > This problem would have got worse when slub came along doing its stupid
>> > unnecessary high-order allocations.
>> >
>> > Billions of years ago a direct-reclaimer had a one-deep cache in the
>> > task_struct into which it freed the page to prevent it from getting
>> > stolen.
>> >
>> > Later, we took that out because pages were being freed into the
>> > per-cpu-pages magazine, which is effectively task-local anyway. __But
>> > per-cpu-pages are only for order-0 pages. __See slub stupidity, above.
>> >
>> > I expect that this is happening so repeatably because the
>> > direct-reclaimer is dong a sleep somewhere after freeing the pages it
>> > needs - if it wasn't doing that then surely the window wouldn't be wide
>> > enough for it to happen so often. __But I didn't look.
>> >
>> > Suitable fixes might be
>> >
>> > a) don't go to sleep after the successful direct-reclaim.
>>
>> It can't make sure success since direct reclaim needs sleep with !GFP_AOMIC.
>
> It doesn't necessarily need to sleep *after* successfully freeing
> pages.  If it needs to sleep then do it before or during the freeing.
>
>> >
>> > b) reinstate the one-deep task-local free page cache.
>>
>> I like b) so how about this?
>> Just for the concept.
>>
>> @@ -1880,7 +1881,7 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask,
>> unsigned int order,
>>         reclaim_state.reclaimed_slab = 0;
>>         p->reclaim_state = &reclaim_state;
>>
>> -       *did_some_progress = try_to_free_pages(zonelist, order,
>> gfp_mask, nodemask);
>> +       *did_some_progress = try_to_free_pages(zonelist, order,
>> gfp_mask, nodemask, &ret_pages);
>>
>>         p->reclaim_state = NULL;
>>         lockdep_clear_current_reclaim_state();
>> @@ -1892,10 +1893,11 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask,
>> unsigned int order,
>>                 return NULL;
>>
>>  retry:
>> -       page = get_page_from_freelist(gfp_mask, nodemask, order,
>> -                                       zonelist, high_zoneidx,
>> -                                       alloc_flags, preferred_zone,
>> -                                       migratetype);
>> +       if(!list_empty(&ret_pages)) {
>> +               page = lru_to_page(ret_pages);
>> +               list_del(&page->lru);
>> +               free_page_list(&ret_pages);
>> +       }
>
> Maybe.  Or just pass a page*.
>

I did it more detailed but It's not completed and test at all.
Please consider just RFC.

-- CUT_HERE --