On Mon, Dec 19, 2022 at 08:48:10AM +0000, Huang, Kai wrote:

On Mon, 2022-12-19 at 15:53 +0800, Chao Peng wrote:

quoted

quoted

[...]

quoted

+
+	/*
+	 * These pages are currently unmovable so don't place them into
movable
+	 * pageblocks (e.g. CMA and ZONE_MOVABLE).
+	 */
+	mapping = memfd->f_mapping;
+	mapping_set_unevictable(mapping);
+	mapping_set_gfp_mask(mapping,
+			     mapping_gfp_mask(mapping) & ~__GFP_MOVABLE);

But, IIUC removing __GFP_MOVABLE flag here only makes page allocation from
non-
movable zones, but doesn't necessarily prevent page from being migrated.  My
first glance is you need to implement either a_ops->migrate_folio() or just
get_page() after faulting in the page to prevent.

The current api restrictedmem_get_page() already does this, after the
caller calling it, it holds a reference to the page. The caller then
decides when to call put_page() appropriately.

I tried to dig some history. Perhaps I am missing something, but it seems Kirill
said in v9 that this code doesn't prevent page migration, and we need to
increase page refcount in restrictedmem_get_page():

https://lore.kernel.org/linux-mm/20221129112139.usp6dqhbih47qpjl@box.shutemov.name/ (local)

But looking at this series it seems restrictedmem_get_page() in this v10 is
identical to the one in v9 (except v10 uses 'folio' instead of 'page')?

restrictedmem_get_page() increases page refcount several versions ago so
no change in v10 is needed. You probably missed my reply:

https://lore.kernel.org/linux-mm/20221129135844.GA902164@chaop.bj.intel.com/ (local)

The current solution is clear: unless we have better approach, we will
let restrictedmem user (KVM in this case) to hold the refcount to
prevent page migration.

Thanks,
Chao

Anyway if this is not fixed, then it should be fixed.  Otherwise, a comment at
the place where page refcount is increased will be helpful to help people
understand page migration is actually prevented.