mm/rmap: Should vma_migratable be checked in try_to_migrate_one?

mm/rmap: Should vma_migratable be checked in try_to_migrate_one?

[Riwan]

[-- Attachment #1: Type: text/plain, Size: 1513 bytes --]

Hi, I’m a phd student working on memory management and I have the following question:
Is it intentional to not check for `vma_migratable` in `try_to_migrate_one` ?
If the vma is not migratable, it would be logical to cancel the page migration at this point.

From what I understand, `vma_migratable` is called to check if a pages in a specific VMA can be migrated or not.
From what I can see, user-facing migration paths (mbind, move_pages) and NUMA balancing already check wether migration is possible or not with `vma_migratable`, but other migration path, such as compaction with kcompactd, ignore this check.
As `try_to_migrate_one` is the only place where we have access to the vma (due to reverse mapping), it would be logical to check here for `vma_migratable` (before we only have access to the folio).

I’m still new to the Linux Kernel, and may have missed critical subsystem / behavior, and I would be curious to know if this was overlooked or if there is a reason behind not calling `vma_migratable`.

From what I could get, here is the relevant context:
- Implem of vma_migratable : mm/mempolicy: check hugepage migration is supported by arch in vma_migratable() (https://lore.kernel.org/all/20200402041052.Y4zsoYman%25akpm@linux-foundation.org/)
- Implem of try_to_migrate_one, previously managed in try_to_unmap_one : mm/rmap: split migration into its own function (https://lore.kernel.org/all/20210701015416.0t4MkxtDR%25akpm@linux-foundation.org/)

Thanks a lot

Riwan Coëffic

[-- Attachment #2: Type: text/html, Size: 2004 bytes --]

Re: mm/rmap: Should vma_migratable be checked in try_to_migrate_one?

[Gregory Price]

On Wed, Mar 11, 2026 at 02:46:55PM +0100, Riwan wrote:
> Hi, I’m a phd student working on memory management and I have the following question:
> Is it intentional to not check for `vma_migratable` in `try_to_migrate_one` ?
> If the vma is not migratable, it would be logical to cancel the page migration at this point.
> 
> From what I understand, `vma_migratable` is called to check if a pages in a specific VMA can be migrated or not.
> From what I can see, user-facing migration paths (mbind, move_pages) and NUMA balancing already check wether migration is possible or not with `vma_migratable`, but other migration path, such as compaction with kcompactd, ignore this check.
> As `try_to_migrate_one` is the only place where we have access to the vma (due to reverse mapping), it would be logical to check here for `vma_migratable` (before we only have access to the folio).
> 
> I’m still new to the Linux Kernel, and may have missed critical subsystem / behavior, and I would be curious to know if this was overlooked or if there is a reason behind not calling `vma_migratable`.
> 
> From what I could get, here is the relevant context:
> - Implem of vma_migratable : mm/mempolicy: check hugepage migration is supported by arch in vma_migratable() (https://lore.kernel.org/all/20200402041052.Y4zsoYman%25akpm@linux-foundation.org/)
> - Implem of try_to_migrate_one, previously managed in try_to_unmap_one : mm/rmap: split migration into its own function (https://lore.kernel.org/all/20210701015416.0t4MkxtDR%25akpm@linux-foundation.org/)
> 

Cursory look, it appears that vma_migratable is intended as a shortcut
to determine if an entire given VMA is otherwise not eligible for
migration (if it's backed by dax, then it's assumed the data should
just live there always and forever).

This is different than kcompact and others which touch individual
folios - but will have different ways to determine the same thing.

If you're just trying to migrate one folio, you can get everything you
need mostly from the folio's flags, and so most of the callers you're
concerned about most likely do eactly this.

~Gregory