Re: [PATCH v3 3/8] mm/hugetlb: unify migration callbacks

[Vlastimil Babka]

On 6/26/20 6:02 AM, Joonsoo Kim wrote:
> 2020ë 6ì 25ì (ë) ìí 8:26, Michal Hocko <mhocko@xxxxxxxxxx>ëì ìì:
>>
>> On Tue 23-06-20 15:13:43, Joonsoo Kim wrote:
>> > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
>> >
>> > There is no difference between two migration callback functions,
>> > alloc_huge_page_node() and alloc_huge_page_nodemask(), except
>> > __GFP_THISNODE handling. This patch adds an argument, gfp_mask, on
>> > alloc_huge_page_nodemask() and replace the callsite for
>> > alloc_huge_page_node() with the call to
>> > alloc_huge_page_nodemask(..., __GFP_THISNODE).
>> >
>> > It's safe to remove a node id check in alloc_huge_page_node() since
>> > there is no caller passing NUMA_NO_NODE as a node id.
>>
>> Yes this is indeed safe. alloc_huge_page_node used to be called from
>> other internal hugetlb allocation layer and that allowed NUMA_NO_NODE as
>> well. Now it is called only from the mempolicy migration callback and
>> that always specifies a node and want to stick with that node.
>>
>> But I have to say I really dislike the gfp semantic because it is
>> different from any other allocation function I can think of. It
>> specifies what to be added rather than what should be used.
>>
>> Removing the function is ok but please use the full gfp mask instead
>> or if that is impractical for some reason (wich shouldn't be the case
>> as htlb_alloc_mask should be trivial to make static inline) make it
>> explicit that this is not a gfp_mask but a gfp modifier and explicitly
>> state which modifiers are allowed.
> 
> Okay. I will try to solve your concern. Concrete solution is not yet prepared
> but perhaps I will use full gfp_mask by using htlb_alloc_mask() in caller sites.

Yeah, that should be feasible. alloc_huge_page_vma() already does
htlb_alloc_mask(h). In alloc_new_node_page() and new_page_nodemask() it would be
consistent with the other cases handled there (THP and base).

> Thanks.
>