Re: [mm 4.15-rc8] Random oopses under memory pressure.

[Dave Hansen]

On 01/18/2018 06:45 AM, Kirill A. Shutemov wrote:
> On Thu, Jan 18, 2018 at 06:38:10AM -0800, Dave Hansen wrote:
>> On 01/18/2018 05:12 AM, Kirill A. Shutemov wrote:
>>> -		if (pte_page(*pvmw->pte) - pvmw->page >=
>>> -				hpage_nr_pages(pvmw->page)) {
>> Is ->pte guaranteed to map a page which is within the same section as
>> pvmw->page?  Otherwise, with sparsemem (non-vmemmap), the pointer
>> arithmetic won't work.
> No, it's not guaranteed. It can be arbitrary page.
> 
> The arithmetic won't work because they are different "memory objects"?

No, because sections' mem_map[]s can be allocated non-contiguously.
Section 1's might be a lower virtual address than Section 0's.

They're allocated not unlike this:

	mem_section[0]->section_mem_map = kmalloc(SECTION_SIZE);
	mem_section[1]->section_mem_map = kmalloc(SECTION_SIZE);
	...

The first pfn in section 1 and the last pfn in section 0 are adjacent
PFNs, but their 'struct page' might have virtual addresses that are TB
apart.