Hi Mel,
On 02/18/2013 11:17 PM, Mel Gorman wrote:
>>> > > <SNIP>
>>> > >
>>> > > result. It's a little clumsy but the memory hot-remove failure message
>>> > > could list what applications have pinned the pages that cannot be
>>> > > removed
>>> > > so the administrator has the option of force-killing the application. It
>>> > > is possible to discover what application is pinning a page from
>>> > > userspace
>>> > > but it would involve an expensive search with /proc/kpagemap
>>> > >
>>>>> > >>> + if (migrate_pre_flag && !isolate_err) {
>>>>> > >>> + ret = migrate_pages(&pagelist, alloc_migrate_target, 1,
>>>>> > >>> + false, MIGRATE_SYNC,
>>>>> > >>> MR_SYSCALL);
>>> > >
>>> > > The conversion of alloc_migrate_target is a bit problematic. It strips
>>> > > the __GFP_MOVABLE flag and the consequence of this is that it converts
>>> > > those allocation requests to MIGRATE_UNMOVABLE. This potentially is a
>>> > > large
>>> > > number of pages, particularly if the number of
>>> > > get_user_pages_non_movable()
>>> > > increases for short-lived pins like direct IO.
>> >
>> > Sorry, I don't quite understand here neither. If we use the following new
>> > migration allocation function as you said, the increasing number of
>> > get_user_pages_non_movable() will also lead to large numbers of
>> > MIGRATE_UNMOVABLE
>> > pages. What's the difference, do I miss something?
>> >
> The replacement function preserves the __GFP_MOVABLE flag. It cannot use
> ZONE_MOVABLE but otherwise the newly allocated page will be grouped with
> other movable pages.
Ah, got it " But GFP_MOVABLE is not only a zone specifier but also an
allocation policy.".
Could I clear __GFP_HIGHMEM flag in alloc_migrate_target depending on private
parameter so
that we can keep MIGRATE_UNMOVABLE policy also allocate page none movable zones
with little
change?
Does that approach work? Otherwise I have to follow your suggestion.
thanks,
linfeng
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
