On Wed, May 27, 2020 at 08:46:22PM -0700, Hugh Dickins wrote:

Hi Jerome,

On Wed, 27 May 2020, Jerome Glisse wrote:

quoted

So any arch code which uses page_mapping_file() might get the wrong
answer, this function will return NULL for a swap backed page which
can be a shmem pages. But shmem pages can still be shared among
multiple process (and possibly at different virtual addresses if
mremap was use).

Attached is a patch that changes page_mapping_file() to return the
shmem mapping for swap backed shmem page. I have not tested it (no
way for me to test all those architecture) and i spotted this while
working on something else. So i hope someone can take a closer look.

I'm certainly no expert on flush_dcache_page() and friends, but I'd
be very surprised if such a problem exists, yet has gone unnoticed
for so long.  page_mapping_file() itself is fairly new, added when
a risk of crashing on a race with swapoff came in: but the previous
use of page_mapping() would have suffered equally if there were such
a cache flushinhg problem here.

And I'm afraid your patch won't do anything to help if there is a
problem: very soon after shmem calls add_to_swap_cache(), it calls
shmem_delete_from_page_cache(), which sets page->mapping to NULL.

But I can assure you that a shmem page (unlike an anon page) is never
put into swap cache while it is mapped into userspace, and never
mapped into userspace while it is still in swap cache: does that help?

You are right i missed/forgot the part where shmem is never swapcache
and mapped at the same time, thus page_mapping_file() can return NULL
for those as they can no longer have alias mapping.

Thank you Hugh
Jérôme


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel