Re: [PATCH v4 14/20] mm/rmap: use anon pgoff to track MAP_PRIVATE file-backed anon folios

On 8/6/26 22:21, Lorenzo Stoakes (ARM) wrote:

Currently anonymous folios belonging to CoW'd MAP_PRIVATE file-backed
mappings are indexed by their page offset within the file in which they
were originally mapped.

This differs from anonymous folios belonging to pure anon mappings which
are indexed by their anonymous page offset (the address at which they'd
belong in the VMA when first faulted).

This change fixes this inconsistency, always indexing anonymous folios by
their anonymous page offset regardless of the VMA to which they belong.

The foundations have been laid such that we need only switch this
functionality on such by:

* Using linear_anon_page_index() in __folio_set_anon() to assign the
  folio's index to the anonymous linear index rather than the file-backed
  one.

* Otherwise using linear_anon_page_index() in all instances where
  anonymous folios are being referenced or manipulated.

* Replacing vma_address() with vma_filebacked_address() or
  vma_anon_address() as appropriate.

* Updating the merging logic to check that anonymous page offsets are
  aligned as well as filebacked ones for MAP_PRIVATE file-backed VMAs,
  introducing needs_adjacent_anon_pgoff() to figure out when this is
  required.

* Updating linear_folio_page_index() to invoke linear_anon_page_index()
  if the folio is anonymous.

* Updating vma_address_end() to use the VMA's anonymous page offset when
  pvmw->pgoff is anonymous.

* Correcting folio_within_range() to use anonymous page offset for
  anonymous folios.

This will have no impact on merging of anonymous VMAs, whose page offset
and anonymous page offset are identical, nor will it impact shared
file-backed VMAs, which will continue to be merged based on the file-backed
page offset.

However, MAP_PRIVATE file-backed mappings must now be aligned on anonymous
page offset as well.

In most instances this should have no impact on merging of file-backed
mappings, which are usually not merged all that often, let alone
MAP_PRIVATE mapped ones, and rarely remapped and faulted before being moved
back in place (the case in which a merge may now fail).

One subtle impact of this change is in NUMA interleaving - since commit
88c91dc58582 ("mempolicy: migration attempt to match interleave nodes"),
migration heuristically tries to maintain interleaving behaviour matching
the policy using folio indices.

When doing migration of CoW'd MAP_PRIVATE-file backed ranges, the 'base'
upon which the interleaving behaviour is performed will vary for these
ranges. However the commit notes that ranges spanning multiple VMAs will
already cause varying bases, and that this is an acceptable approximation.

It is very unlikely real world use-cases will be impacted by
this (MAP_PRIVATE file-backed mappings are already an edge case), and all
that will happen is that such ranges will cause interleaving to be rotated
over the CoW'd range, with little to no impact.

This commit lays the foundations for future scalable CoW work which needs
to track some remaps, meaning that most remap tracking can be avoided, and
in nearly all cases the anonymous page offset will be able to be used to
quickly find the VMA in an mm.

Note that the need_rmap_locks check doesn't need to be updated, as any
remapping will offset both the anonymous and file-backed page offset, so it
suffices to check only one.

Signed-off-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
---

Acked-by: David Hildenbrand (Arm) <david@kernel.org>

-- 
Cheers,

David