Re: [PATCH 2/3] mm/sparsemem: get physical address to page struct instead of virtual address to pfn

[Baoquan He]

Hi Dan,

On 02/06/20 at 06:19pm, Dan Williams wrote:
> On Thu, Feb 6, 2020 at 3:17 PM Wei Yang <richardw.yang@xxxxxxxxxxxxxxx> wrote:
> > diff --git a/mm/sparse.c b/mm/sparse.c
> > index b5da121bdd6e..56816f653588 100644
> > --- a/mm/sparse.c
> > +++ b/mm/sparse.c
> > @@ -888,7 +888,7 @@ int __meminit sparse_add_section(int nid, unsigned long start_pfn,
> >         /* Align memmap to section boundary in the subsection case */
> >         if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP) &&
> >                 section_nr_to_pfn(section_nr) != start_pfn)
> > -               memmap = pfn_to_kaddr(section_nr_to_pfn(section_nr));
> > +               memmap = pfn_to_page(section_nr_to_pfn(section_nr));
> 
> Yes, this looks obviously correct. This might be tripping up
> makedumpfile. Do you see any practical effects of this bug? The kernel
> mostly avoids ->section_mem_map in the vmemmap case and in the
> !vmemmap case section_nr_to_pfn(section_nr) should always equal
> start_pfn.

The practical effects is that the memmap for the first unaligned section will be lost
when destroy namespace to hot remove it. Because we encode the ->section_mem_map
into mem_section, and get memmap from the related mem_section to free it in
section_deactivate(). In fact in vmemmap, we don't need to encode the ->section_mem_map
with memmap.

By the way, sub-section support is only valid in vmemmap case, right?
Seems yes from code, but I don't find any document to prove it.

Thanks
Baoquan