Re: [PATCH 1/5] mm, compaction: pass classzone_idx and alloc_flags to watermark checking

[Vlastimil Babka]

On 10/28/2014 08:16 AM, Joonsoo Kim wrote:> On Mon, Oct 27, 2014 at 
10:11:31AM +0100, Vlastimil Babka wrote:
>> On 10/27/2014 07:46 AM, Joonsoo Kim wrote:
>>> On Tue, Oct 07, 2014 at 05:33:35PM +0200, Vlastimil Babka wrote:
>>>
>>> Hello,
>>>
>>> compaction_suitable() has one more zone_watermark_ok(). Why is it
>>> unchanged?
>>
>> Hi,
>>
>> it's a check whether there are enough free pages to perform compaction,
>> which means enough migration targets and temporary copies during
>> migration. These allocations are not affected by the flags of the
>> process that makes the high-order allocation.
>
> Hmm...
>
> To check whether enough free page is there or not needs zone index and
> alloc flag. What we need to ignore is just order information, IMO.
> If there is not enough free page in that zone, compaction progress
> doesn't have any meaning. It will fail due to shortage of free page
> after successful compaction.

I thought that the second check in compaction_suitable() makes sure of 
this, but now I see it's in fact not.
But i'm not sure if we should just put the flags in the first check, as 
IMHO the flags should only affect the final high-order allocation, not 
also the temporary pages needed for migration?

BTW now I'm not even sure that the 2UL << order part makes sense 
anymore. The number of pages migrated at once is always restricted by 
COMPACT_CLUSTER_MAX, so why would we need more than that to cover migration?
Also the order of checks seems wrong. It should return COMPACT_PARTIAL 
"If the allocation would succeed without compaction" but that only can 
happen after passing the check if the zone has the extra 1UL << order 
for migration. Do you agree?

> I guess that __isolate_free_page() is also good candidate to need this
> information in order to prevent compaction from isolating too many
> freepage in low memory condition.

I don't see how it would help here. It's temporary allocations for page 
migration. How would passing classzone_idx and alloc_flags prevent 
isolating too many?

> Thanks.
>

--
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/