Re: [PATCH 7/7] mm/hotplug: fix hot remove failure in SPARSEMEM|!VMEMMAP case

[Wei Yang]

On Mon, Feb 10, 2020 at 03:54:06PM +0800, Baoquan He wrote:
>On 02/10/20 at 02:08pm, Wei Yang wrote:
>> On Mon, Feb 10, 2020 at 11:41:05AM +0800, Baoquan He wrote:
>> >On 02/10/20 at 07:52am, Wei Yang wrote:
>> >> >---
>> >> > mm/sparse.c | 4 +++-
>> >> > 1 file changed, 3 insertions(+), 1 deletion(-)
>> >> >
>> >> >diff --git a/mm/sparse.c b/mm/sparse.c
>> >> >index 623755e88255..345d065ef6ce 100644
>> >> >--- a/mm/sparse.c
>> >> >+++ b/mm/sparse.c
>> >> >@@ -854,13 +854,15 @@ static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
>> >> > 			ms->usage = NULL;
>> >> > 		}
>> >> > 		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
>> >> >-		ms->section_mem_map = (unsigned long)NULL;
>> >> > 	}
>> >> > 
>> >> > 	if (section_is_early && memmap)
>> >> > 		free_map_bootmem(memmap);
>> >> > 	else
>> >> > 		depopulate_section_memmap(pfn, nr_pages, altmap);
>> >> 
>> >> The crash happens in depopulate_section_memmap() when trying to get memmap by
>> >> pfn_to_page(). Can we pass memmap directly?
>> >
>> >Yes, that's also a good idea. While it needs to add a parameter for
>> >depopulate_section_memmap(), the parameter is useless for VMEMMAP
>> >though, I personally prefer the current fix which is a little simpler.
>> >
>> 
>> Not a new parameter, but replace pfn with memmap.
>> 
>> Not sure why the parameter is useless for VMEMMAP? memmap will be assigned to
>> start and finally pass to vmemmap_free().
>
>In section_deactivate(), per the code comments from Dan, we can know
>that:
>
>	/*
>	 * section which only contains bootmem will be handled by
>	 * free_map_bootmem(), including a complete section, or partial
>	 * section which only has memory starting from the begining.
>	 */
>        if (section_is_early && memmap)
>                free_map_bootmem(memmap);                                                                                                         
>        else
>	/* 
>	 * section which contains region mixing bootmem with hot added
>	 * sub-section region, only sub-section region, complete
>	 * section. And in the mxied case, if hot remove the hot added
>	 * sub-section aligned part, no memmap is got in the current
>	 * code. So we still need pfn to calculate it for vmemmap case.
>	 * To me, whenever we need, it looks better that we always use
>	 * pfn to get its own memmap. 
>	 */
>                depopulate_section_memmap(pfn, nr_pages, altmap);
>
>This is why I would like to keep the current logic as is,only one line
>of code adjusting can fix the issue. Please let me know if I miss
>anything.
>

You are right. I missed this point.

>
>> 
>> >Anyway, both is fine to me, I can update if you think passing memmap is
>> >better.
>> >
>> >> 
>> >> >+
>> >> >+	if(!rc)
>> >> >+		ms->section_mem_map = (unsigned long)NULL;
>> >> > }
>> >> > 
>> >> > static struct page * __meminit section_activate(int nid, unsigned long pfn,
>> >> >-- 
>> >> >2.17.2
>> >> 
>> >> -- 
>> >> Wei Yang
>> >> Help you, Help me
>> >> 
>> 
>> -- 
>> Wei Yang
>> Help you, Help me
>> 

-- 
Wei Yang
Help you, Help me