Re: compaction: VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn))

[Qian Cai]

> On May 5, 2020, at 8:43 AM, Baoquan He <bhe@xxxxxxxxxx> wrote:
> 
> Hi,
> 
> On 04/24/20 at 09:45am, Qian Cai wrote:
>> 
>> 
>>> On Apr 23, 2020, at 11:43 PM, Baoquan He <bhe@xxxxxxxxxx> wrote:
>>> 
>>> On 04/23/20 at 05:25pm, Qian Cai wrote:
>>>> Compaction starts to crash below on linux-next today. The faulty page belongs to Node 0 DMA32 zone.
>>>> Iâll continue to narrow it down, but just want to give a headup in case someone could beat me to it.
>>>> 
>>>> Debug output from free_area_init_core()
>>>> [    0.000000] KK start page = ffffea0000000040, end page = ffffea0000040000, nid = 0 DMA
>>>> [    0.000000] KK start page = ffffea0000040000, end page = ffffea0004000000, nid = 0 DMA32
>>>> [    0.000000] KK start page = ffffea0004000000, end page = ffffea0012000000, nid = 0 NORMAL
>>>> [    0.000000] KK start page = ffffea0012000000, end page = ffffea0021fc0000, nid = 4 NORMAL
>>> 
>>> Where are these printed? They are the direct mapping address of page?
>> 
>> From this debug patch. Yes, direct mapping.
> 
> Can you try below patch? I may get why this is caused, not sure if the
> place is right. 
> 
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 177c11a8f3b9..e26972f26414 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -1409,7 +1409,9 @@ fast_isolate_freepages(struct compact_control *cc)
> 				cc->free_pfn = highest;
> 			} else {
> 				if (cc->direct_compaction && pfn_valid(min_pfn)) {
> -					page = pfn_to_page(min_pfn);
> +					page = pageblock_pfn_to_page(min_pfn,
> +						pageblock_end_pfn(min_pfn),
> +						cc->zone);
> 					cc->free_pfn = min_pfn;
> 				}
> 			}

I have not had luck to reproduce this again yet, but feel free to move forward with the patch anyway if you are comfortable to do so, so at least people could review it properly.