Re: [PATCH v6 08/14] mm/gup: do not migrate zero page

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Jan 19, 2021 at 08:43:27PM -0500, Pavel Tatashin wrote:
> On some platforms ZERO_PAGE(0) might end-up in a movable zone. Do not
> migrate zero page in gup during longterm pinning as migration of zero page
> is not allowed.
> 
> For example, in x86 QEMU with 16G of memory and kernelcore=5G parameter, I
> see the following:
> 
> Boot#1: zero_pfn  0x48a8d zero_pfn zone: ZONE_DMA32
> Boot#2: zero_pfn 0x20168d zero_pfn zone: ZONE_MOVABLE
> 
> On x86, empty_zero_page is declared in .bss and depending on the loader
> may end up in different physical locations during boots.
> 
> Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com>
>  include/linux/mmzone.h | 4 ++++
>  mm/gup.c               | 2 ++
>  2 files changed, 6 insertions(+)
> 
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index fc99e9241846..f67427a8f22b 100644
> +++ b/include/linux/mmzone.h
> @@ -427,6 +427,10 @@ enum zone_type {
>  	 *    techniques might use alloc_contig_range() to hide previously
>  	 *    exposed pages from the buddy again (e.g., to implement some sort
>  	 *    of memory unplug in virtio-mem).
> +	 * 6. ZERO_PAGE(0), kernelcore/movablecore setups might create
> +	 *    situations where ZERO_PAGE(0) which is allocated differently
> +	 *    on different platforms may end up in a movable zone. ZERO_PAGE(0)
> +	 *    cannot be migrated.
>  	 *
>  	 * In general, no unmovable allocations that degrade memory offlining
>  	 * should end up in ZONE_MOVABLE. Allocators (like alloc_contig_range())
> diff --git a/mm/gup.c b/mm/gup.c
> index 857b273e32ac..fdd5cda30a07 100644
> +++ b/mm/gup.c
> @@ -1580,6 +1580,8 @@ static long check_and_migrate_cma_pages(struct mm_struct *mm,
>  		 * of the CMA zone if possible.
>  		 */
>  		if (is_migrate_cma_page(head)) {
> +			if (is_zero_pfn(page_to_pfn(head)))
> +				continue;

I think you should put this logic in is_pinnable_page()

Jason