Re: [PATCH] mm/compaction: fix low_pfn advance on isolating hugetlb

[Wei Yang <richard.weiyang@gmail.com>]

On Thu, Sep 11, 2025 at 12:28:24PM -0400, Zi Yan wrote:
>On 11 Sep 2025, at 2:30, Wei Yang wrote:
>
>> On Thu, Sep 11, 2025 at 03:34:13AM +0000, Wei Yang wrote:
>>> On Wed, Sep 10, 2025 at 09:50:09PM -0400, Zi Yan wrote:
>>>> On 10 Sep 2025, at 21:38, Zi Yan wrote:
>>>>
>>>>> On 10 Sep 2025, at 21:35, Zi Yan wrote:
>>>>>
>>>>>> On 10 Sep 2025, at 21:25, Wei Yang wrote:
>>>>>>
>>>>>>> On Wed, Sep 10, 2025 at 09:22:40AM +0000, Wei Yang wrote:
>>>>>>>> Commit 56ae0bb349b4 ("mm: compaction: convert to use a folio in
>>>>>>>> isolate_migratepages_block()") converts api from page to folio. But the
>>>>>>>> low_pfn advance for hugetlb page seems wrong when low_pfn doesn't point
>>>>>>>> to head page.
>>>>>>>>
>>>>>>>> Originally, if page is a hugetlb tail page, compound_nr() return 1,
>>>>>>>> which means low_pfn only advance one in next iteration. After the
>>>>>>>> change, low_pfn would advance more than the hugetlb range, since
>>>>>>>> folio_nr_pages() always return total number of the large page. This
>>>>>>>> results in skipping some range to isolate and then to migrate.
>>>>>>>>
>>>>>>>> The worst case for alloc_contig is it does all the isolation and
>>>>>>>> migration, but finally find some range is still not isolated. And then
>>>>>>>> undo all the work and try a new range.
>>>>>>>>
>>>>>>>> Advance low_pfn to the end of hugetlb.
>>>>>>>>
>>>>>>>> Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
>>>>>>>> Fixes: 56ae0bb349b4 ("mm: compaction: convert to use a folio in isolate_migratepages_block()")
>>>>
>>>> This behavior seems to be introduced by commit 369fa227c219 ("mm: make
>>>> alloc_contig_range handle free hugetlb pages”). The related change is:
>>>>
>>>> +               if (PageHuge(page) && cc->alloc_contig) {
>>>> +                       ret = isolate_or_dissolve_huge_page(page);
>>>> +
>>>> +                       /*
>>>> +                        * Fail isolation in case isolate_or_dissolve_huge_page()
>>>> +                        * reports an error. In case of -ENOMEM, abort right away.
>>>> +                        */
>>>> +                       if (ret < 0) {
>>>> +                                /* Do not report -EBUSY down the chain */
>>>> +                               if (ret == -EBUSY)
>>>> +                                       ret = 0;
>>>> +                               low_pfn += (1UL << compound_order(page)) - 1;
>>>
>>> The compound_order(page) return 1 for a tail page.
>>>
>>> See below.
>>>
>>>> +                               goto isolate_fail;
>>>> +                       }
>>>> +
>>>> +                       /*
>>>> +                        * Ok, the hugepage was dissolved. Now these pages are
>>>> +                        * Buddy and cannot be re-allocated because they are
>>>> +                        * isolated. Fall-through as the check below handles
>>>> +                        * Buddy pages.
>>>> +                        */
>>>> +               }
>>>> +
>>>>
>>>>>>>> Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
>>>>>>>> Cc: Oscar Salvador <osalvador@suse.de>
>>>>>>>
>>>>>>> Forgot to cc stable.
>>>>>>>
>>>>>>> Cc: <stable@vger.kernel.org>
>>>>>>
>>>>>> Is there any bug report to justify the backport? Since it is more likely
>>>>>> to be a performance issue instead of a correctness issue.
>>>>>>
>>>>>>>
>>>>>>>> ---
>>>>>>>> mm/compaction.c | 2 +-
>>>>>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/mm/compaction.c b/mm/compaction.c
>>>>>>>> index bf021b31c7ec..1e8f8eca318c 100644
>>>>>>>> --- a/mm/compaction.c
>>>>>>>> +++ b/mm/compaction.c
>>>>>>>> @@ -989,7 +989,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
>>>>>>>> 				 * Hugepage was successfully isolated and placed
>>>>>>>> 				 * on the cc->migratepages list.
>>>>>>>> 				 */
>>>>>>>> -				low_pfn += folio_nr_pages(folio) - 1;
>>>>>>>> +				low_pfn += folio_nr_pages(folio) - folio_page_idx(folio, page) - 1;
>>>>>>>
>>>>>>> One question is why we advance compound_nr() in original version.
>>>>>>>
>>>>>>> Yes, there are several places advancing compound_nr(), but it seems to iterate
>>>>>>> on the same large page and do the same thing and advance 1 again.
>>>>>>>
>>>>>>> Not sure which part story I missed.
>>>>>>
>>>>>> isolate_migratepages_block() starts from the beginning of a pageblock.
>>>>>> How likely the code hit in the middle of a hugetlb?
>>>>>>
>>>>>
>>>>> In addition, there are two other “low_pfn += (1UL << order) - 1”
>>>>> in the if (PageHuge(page)), why not change them too if you think
>>>>> page can point to the middle of a hugetlb?
>>>>>
>>>
>>> The order here is get from compound_order(page), which is 1 for a tail page.
>>>
>>> So it looks right. Maybe I misunderstand it?
>>
>> Oops, compound_order(page) returns 0 for tail page.
>>
>> What I want to say is low_pfn advance by 1 for tail page. Sorry for the
>> misleading.
>
>OK, that sounds inefficient and inconsistent with your fix.
>
>While at it, can you also change two “low_pfn += (1UL << order) - 1” to skip
>the rest of hugetlb?
>

Sure, glad to.

You prefer do the fix in one patch or have a separate one?

>--
>Best Regards,
>Yan, Zi

-- 
Wei Yang
Help you, Help me