Re: [PATCH] mm, page_alloc: fix build_zonerefs_node()

[David Hildenbrand <david@redhat.com>]

On 07.04.22 14:04, Michal Hocko wrote:
> On Thu 07-04-22 13:58:44, David Hildenbrand wrote:
> [...]
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index 3589febc6d31..130a2feceddc 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -6112,10 +6112,8 @@ static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
>>>  	do {
>>>  		zone_type--;
>>>  		zone = pgdat->node_zones + zone_type;
>>> -		if (managed_zone(zone)) {
>>> -			zoneref_set_zone(zone, &zonerefs[nr_zones++]);
>>> -			check_highest_zone(zone_type);
>>> -		}
>>> +		zoneref_set_zone(zone, &zonerefs[nr_zones++]);
>>> +		check_highest_zone(zone_type);
>>>  	} while (zone_type);
>>>  
>>>  	return nr_zones;
>>
>> I don't think having !populated zones in the zonelist is a particularly
>> good idea. Populated vs !populated changes only during page
>> onlininge/offlining.
>>
>> If I'm not wrong, with your patch we'd even include ZONE_DEVICE here ...
> 
> What kind of problem that would cause? The allocator wouldn't see any
> pages at all so it would fallback to the next one. Maybe kswapd would
> need some tweak to have a bail out condition but as mentioned in the
> thread already. !populated or !managed for that matter are not all that
> much different from completely depleted zones. The fact that we are
> making that distinction has led to some bugs and I suspect it makes the
> code more complex without a very good reason.

I assume performance problems. Assume you have an ordinary system with
multiple NUMA nodes and no MOVABLE memory. Most nodes will only have
ZONE_NORMAL. Yet, you'd include ZONE_DMA* and ZONE_MOVABLE that will
always remain empty to be traversed on each and every allocation
fallback. Of course, we could measure, but IMHO at least *that* part of
memory onlining/offlining is not the complicated part :D

Populated vs. !populated is under pretty good control via page
onlining/offlining. We have to be careful with "managed pages", because
that's a moving target, especially with memory ballooning. And I assume
that's the bigger source of bugs.

> 
>> I'd vote for going with the simple fix first, which should be good
>> enough AFAIKT.
> 
> yes, see the other reply
> 

I think we were composing almost simultaneously :)

-- 
Thanks,

David / dhildenb