Re: [PATCH v4 4/5] mm: Fix page reference leak in soft_offline_page()

[Dan Williams]

On Wed, Jan 13, 2021 at 5:50 PM HORIGUCHI NAOYA(堀口　直也)
<naoya.horiguchi@xxxxxxx> wrote:
>
> On Wed, Jan 13, 2021 at 04:43:32PM -0800, Dan Williams wrote:
> > The conversion to move pfn_to_online_page() internal to
> > soft_offline_page() missed that the get_user_pages() reference taken by
> > the madvise() path needs to be dropped when pfn_to_online_page() fails.
> > Note the direct sysfs-path to soft_offline_page() does not perform a
> > get_user_pages() lookup.
> >
> > When soft_offline_page() is handed a pfn_valid() &&
> > !pfn_to_online_page() pfn the kernel hangs at dax-device shutdown due to
> > a leaked reference.
> >
> > Fixes: feec24a6139d ("mm, soft-offline: convert parameter to pfn")
> > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > Cc: Naoya Horiguchi <nao.horiguchi@xxxxxxxxx>
> > Cc: Michal Hocko <mhocko@xxxxxxxxxx>
> > Reviewed-by: David Hildenbrand <david@xxxxxxxxxx>
> > Reviewed-by: Oscar Salvador <osalvador@xxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx>
> > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
>
> I'm OK if we don't have any other better approach, but the proposed changes
> make code a little messy, and I feel that get_user_pages() might be the
> right place to fix. Is get_user_pages() expected to return struct page with
> holding refcount for offline valid pages?  I thought that such pages are
> only used by drivers for dax-devices, but that might be wrong. Can I ask for
> a little more explanation from this perspective?

The motivation for ZONE_DEVICE is to allow get_user_pages() for "offline" pfns.

soft_offline_page() wants to only operate on "online" pfns.

get_user_pages() on a dax-mapping returns an "offline" ZONE_DEVICE page.

When pfn_to_online_page() fails the get_user_pages() reference needs
to be dropped.

To be honest I dislike the entire flags based scheme for communicating
the fact that page reference obtained by madvise needs to be dropped
later. I'd rather pass a non-NULL 'struct page *' than redo
pfn_to_page() conversions in the leaf functions, but that's a much
larger change.