On 06/01/2016 11:19 AM, Mel Gorman wrote:
> On Tue, May 31, 2016 at 11:44:24PM +0200, Vlastimil Babka wrote:
>> On 05/30/2016 05:56 PM, Mel Gorman wrote:
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index dba8cfd0b2d6..f2c1e47adc11 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -3232,6 +3232,9 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int 
>>> order,
>>>              * allocations are system rather than user orientated
>>>              */
>>>             ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
>>> +           ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
>>> +                                   ac->high_zoneidx, ac->nodemask);
>>> +           ac->classzone_idx = zonelist_zone_idx(ac->preferred_zoneref);
>>>             page = get_page_from_freelist(gfp_mask, order,
>>>                                             ALLOC_NO_WATERMARKS, ac);
>>>             if (page)
>>>
>>
>> Even if that didn't help for this report, I think it's needed too
>> (except the classzone_idx which doesn't exist anymore?).

But you agree that the hunk above should be merged?

>> And I think the following as well. (the changed comment could be also
>> just deleted).
>>
> 
> Why?
> 
> The comment is fine but I do not see why the recalculation would occur.
> 
> In the original code, the preferred_zoneref for statistics is calculated
> based on either the supplied nodemask or cpuset_current_mems_allowed during
> the initial attempt. It then relies on the cpuset checks in the slowpath
> to encorce mems_allowed but the preferred zone doesn't change.
> 
> With your proposed change, it's possible that the
> preferred_zoneref recalculation points to a zoneref disallowed by
> cpuset_current_mems_sllowed. While it'll be skipped during allocation,
> the statistics will still be against a zone that is potentially outside
> what is allowed.

Hmm that's true and I was ready to agree. But then I noticed  that
gfp_to_alloc_flags() can mask out ALLOC_CPUSET for GFP_ATOMIC. So it's
like a lighter version of the ALLOC_NO_WATERMARKS situation. In that
case it's wrong if we leave ac->preferred_zoneref at a position that has
skipped some zones due to mempolicies?

Reply via email to