On 01/20/2020 05:02 PM, Michal Hocko wrote:
> On Mon 20-01-20 14:51:31, Anshuman Khandual wrote:
>>
>>
>> On 01/16/2020 08:48 PM, Michal Hocko wrote:
>>> On Thu 16-01-20 21:50:34, Li Xinhai wrote:
>>>> On 2020-01-16 at 17:56 Michal Hocko wrote:
>>>>> On Thu 16-01-20 04:11:25, Li Xinhai wrote:
>>>>>> Checking hstate at early phase when isolating page, instead of during
>>>>>> unmap and move phase, to avoid useless isolation.
>>>>>
>>>>> Could you be more specific what you mean by isolation and why does it
>>>>> matter? The patch description should really explain _why_ the change is
>>>>> needed or desirable.
>>>>
>>>> The changelog can be improved:
>>>>
>>>> vma_migratable() is called to check if pages in vma can be migrated
>>>> before go ahead to isolate, unmap and move pages. For hugetlb pages,
>>>> hugepage_migration_supported(struct hstate *h) is one factor which
>>>> decide if migration is supported. In current code, this function is called
>>>> from unmap_and_move_huge_page(), after isolating page has
>>>> completed.
>>>> This patch checks hstate from vma_migratable() and avoids isolating pages
>>>> which are not supported.
>>>
>>> This still explains what but not why this is relevant. If by isolating
>>> pages you mean isolate_lru_page then this really a noop for hugetlb
>>> pages. Or do I still misread your changelog?
>>
>> unmap_and_move_hugepage() aborts migrating a HugeTLB page (from the list)
>> if it's corresponding hstate does not support migration.
>
> But all architectures support all hugeltb sizes unless I am missing
> something. If there is some which doesn't then the changelog should
> mention that. I have already asked for runtime effects with no data
> provided.
You are right that all hugetlb sizes are supported right now whether the
platform defines arch_hugetlb_migration_supported() callback or not. But
in theory an override for the arch callback can deny migration support
for certain huge page sizes.
>
> Just to make it clear. I am not objecting to the patch itself. I am
> objecting to the very vague justification. The changelog doesn't explain
> _why_ do we need to change this. Is it a bug, non-optimal code, pure
> code clean up for a more robust code?
AFAICS this tries to solve the problem like a sub-optimal code. But for
now as there are no real HugeTLB cases for an early bail out, there can
be an argument not to add new cost into via vma_migratable() which will
be called more often for non-HugeTLB VMAs. Probably adding a comment in
the code like this might just be sufficient.
diff --git a/include/linux/mempolicy.h b/include/linux/mempolicy.h
index 5228c62..ca9c343 100644
--- a/include/linux/mempolicy.h
+++ b/include/linux/mempolicy.h
@@ -186,6 +186,13 @@ static inline bool vma_migratable(struct vm_area_struct *vma)
return false;
#ifndef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
+ /*
+ * NOTE: hugepage_migration_supported() should have been called here
+ * for an early bail out in cases where HugeTLB page sizes are not
+ * supported for migration. But for now as there are no such real
+ * cases, hence it is better not to add any additional cost here by
+ * calling hugepage_migration_supported().
+ */
if (vma->vm_flags & VM_HUGETLB)
return false;
#endif
