lkml.org 
[lkml]   [2012]   [Jun]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Accounting problem of MIGRATE_ISOLATED freed page
(6/20/12 3:53 AM), Minchan Kim wrote:
> On 06/20/2012 03:32 PM, KOSAKI Motohiro wrote:
>
>> (6/20/12 2:12 AM), Minchan Kim wrote:
>>>
>>> Hi Aaditya,
>>>
>>> I want to discuss this problem on another thread.
>>>
>>> On 06/19/2012 10:18 PM, Aaditya Kumar wrote:
>>>> On Mon, Jun 18, 2012 at 6:13 AM, Minchan Kim <minchan@kernel.org> wrote:
>>>>> On 06/17/2012 02:48 AM, Aaditya Kumar wrote:
>>>>>
>>>>>> On Fri, Jun 15, 2012 at 12:57 PM, Minchan Kim <minchan@kernel.org> wrote:
>>>>>>
>>>>>>>>
>>>>>>>> pgdat_balanced() doesn't recognized zone. Therefore kswapd may sleep
>>>>>>>> if node has multiple zones. Hm ok, I realized my descriptions was
>>>>>>>> slightly misleading. priority 0 is not needed. bakance_pddat() calls
>>>>>>>> pgdat_balanced()
>>>>>>>> every priority. Most easy case is, movable zone has a lot of free pages and
>>>>>>>> normal zone has no reclaimable page.
>>>>>>>>
>>>>>>>> btw, current pgdat_balanced() logic seems not correct. kswapd should
>>>>>>>> sleep only if every zones have much free pages than high water mark
>>>>>>>> _and_ 25% of present pages in node are free.
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sorry. I can't understand your point.
>>>>>>> Current kswapd doesn't sleep if relevant zones don't have free pages above high watermark.
>>>>>>> It seems I am missing your point.
>>>>>>> Please anybody correct me.
>>>>>>
>>>>>> Since currently direct reclaim is given up based on
>>>>>> zone->all_unreclaimable flag,
>>>>>> so for e.g in one of the scenarios:
>>>>>>
>>>>>> Lets say system has one node with two zones (NORMAL and MOVABLE) and we
>>>>>> hot-remove the all the pages of the MOVABLE zone.
>>>>>>
>>>>>> While migrating pages during memory hot-unplugging, the allocation function
>>>>>> (for new page to which the page in MOVABLE zone would be moved) can end up
>>>>>> looping in direct reclaim path for ever.
>>>>>>
>>>>>> This is so because when most of the pages in the MOVABLE zone have
>>>>>> been migrated,
>>>>>> the zone now contains lots of free memory (basically above low watermark)
>>>>>> BUT all are in MIGRATE_ISOLATE list of the buddy list.
>>>>>>
>>>>>> So kswapd() would not balance this zone as free pages are above low watermark
>>>>>> (but all are in isolate list). So zone->all_unreclaimable flag would
>>>>>> never be set for this zone
>>>>>> and allocation function would end up looping forever. (assuming the
>>>>>> zone NORMAL is
>>>>>> left with no reclaimable memory)
>>>>>>
>>>>>
>>>>>
>>>>> Thanks a lot, Aaditya! Scenario you mentioned makes perfect.
>>>>> But I don't see it's a problem of kswapd.
>>>>
>>>> Hi Kim,
>>>
>>> I like called Minchan rather than Kim
>>> Never mind. :)
>>>
>>>>
>>>> Yes I agree it is not a problem of kswapd.
>>>
>>> Yeb.
>>>
>>>>
>>>>> a5d76b54 made new migration type 'MIGRATE_ISOLATE' which is very irony type because there are many free pages in free list
>>>>> but we can't allocate it. :(
>>>>> It doesn't reflect right NR_FREE_PAGES while many places in the kernel use NR_FREE_PAGES to trigger some operation.
>>>>> Kswapd is just one of them confused.
>>>>> As right fix of this problem, we should fix hot plug code, IMHO which can fix CMA, too.
>>>>>
>>>>> This patch could make inconsistency between NR_FREE_PAGES and SumOf[free_area[order].nr_free]
>>>>
>>>>
>>>> I assume that by the inconsistency you mention above, you mean
>>>> temporary inconsistency.
>>>>
>>>> Sorry, but IMHO as for memory hot plug the main issue with this patch
>>>> is that the inconsistency you mentioned above would NOT be a temporary
>>>> inconsistency.
>>>>
>>>> Every time say 'x' number of page frames are off lined, they will
>>>> introduce a difference of 'x' pages between
>>>> NR_FREE_PAGES and SumOf[free_area[order].nr_free].
>>>> (So for e.g. if we do a frequent offline/online it will make
>>>> NR_FREE_PAGES negative)
>>>>
>>>> This is so because, unset_migratetype_isolate() is called from
>>>> offlining code (to set the migrate type of off lined pages again back
>>>> to MIGRATE_MOVABLE)
>>>> after the pages have been off lined and removed from the buddy list.
>>>> Since the pages for which unset_migratetype_isolate() is called are
>>>> not buddy pages so move_freepages_block() does not move any page, and
>>>> thus introducing a permanent inconsistency.
>>>
>>> Good point. Negative NR_FREE_PAGES is caused by double counting by my patch and __offline_isolated_pages.
>>> I think at first MIGRATE_ISOLATE type freed page shouldn't account as free page.
>>>
>>>>
>>>>> and it could make __zone_watermark_ok confuse so we might need to fix move_freepages_block itself to reflect
>>>>> free_area[order].nr_free exactly.
>>>>>
>>>>> Any thought?
>>>>
>>>> As for fixing move_freepages_block(), At least for memory hot plug,
>>>> the pages stay in MIGRATE_ISOLATE list only for duration
>>>> offline_pages() function,
>>>> I mean only temporarily. Since fixing move_freepages_block() for will
>>>> introduce some overhead, So I am not very sure whether that overhead
>>>> is justified
>>>> for a temporary condition. What do you think?
>>>
>>> Yes. I don't like hurt fast path, either.
>>> How about this? (Passed just compile test :( )
>>> The patch's goal is to NOT increase nr_free and NR_FREE_PAGES about freed page into MIGRATE_ISOLATED.
>>>
>>> This patch hurts high order page free path but I think it's not critical because higher order allocation
>>> is rare than order-0 allocation and we already have done same thing on free_hot_cold_page on order-0 free path
>>> which is more hot.
>>
>> Can't we change zone_water_mark_ok_safe() instead of page allocator? memory hotplug is really rare event.
>
>
> +1
>
> Firstly, I want to make zone_page_state(z, NR_FREE_PAGES) itself more accurately because it is used by
> several places. As I looked over places, I can't find critical places except kswapd forever sleep case.
> So it's a nice idea!
>
> In that case, we need zone->lock whenever zone_watermark_ok_safe is called.
> Most of cases, it's unnecessary and it might hurt alloc/free performance when memory pressure is high.
> But if memory pressure is high, it may be already meaningless alloc/free performance.
> So it does make sense, IMHO.
>
> Please raise your hands if anyone has a concern about this.
>
> barrios@bbox:~/linux-next$ git diff
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index d2a515d..82cc0a2 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1748,16 +1748,38 @@ bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
> zone_page_state(z, NR_FREE_PAGES));
> }
>
> -bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
> +bool zone_watermark_ok_safe(struct zone *z, int alloc_order, unsigned long mark,
> int classzone_idx, int alloc_flags)
> {
> + struct free_area *area;
> + struct list_head *curr;
> + int order;
> + unsigned long flags;
> long free_pages = zone_page_state(z, NR_FREE_PAGES);
>
> if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
> free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
>
> - return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
> - free_pages);
> + /*
> + * Memory hotplug/CMA can isolate freed page into MIGRATE_ISOLATE
> + * so that buddy can't allocate it although they are in free list.
> + */
> + spin_lock_irqsave(&z->lock, flags);
> + for (order = 0; order < MAX_ORDER; order++) {
> + int count = 0;
> + area = &(z->free_area[order]);
> + if (unlikely(!list_empty(&area->free_list[MIGRATE_ISOLATE]))) {
> + list_for_each(curr, &area->free_list[MIGRATE_ISOLATE])
> + count++;
> + free_pages -= (count << order);
> + }
> + }
> + if (free_pages < 0)
> + free_pages = 0;
> + spin_unlock_irqrestore(&z->lock, flags);
> +
> + return __zone_watermark_ok(z, alloc_order, mark,
> + classzone_idx, alloc_flags, free_pages);
> }

number of isolate page block is almost always 0. then if we have such counter,
we almost always can avoid zone->lock. Just idea.





\
 
 \ /
  Last update: 2012-06-21 00:01    [W:0.077 / U:0.196 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site