Re: [PATCH v15 00/11] migrate on fault for device pages

[Mika Penttilä]

On 9/25/26 12:07, Christoph Hellwig wrote:
> Who is going to use this? The series seems to lack an actual user for it.

Two users:

For one, existing migrate_vma_setup() callers, unchanged. The series
reworks migrate_vma_setup() to collect via hmm_range_fault() internally
(it sets HMM_PFN_REQ_MIGRATE and calls migrate_hmm_range_setup(),
mm/migrate_device.c). So every in-tree migrate_vma consumer — nouveau,
amdgpu/amdkfd, drm/xe via drm_pagemap, lib/test_hmm — runs on the new path.

Second, the point is to fault-in and migrate in a single page-table walk
instead of the 2–3 walks drivers do today . It can be driven by
hmm_range_fault() directly instead of ping-ponging between hmm fault and
migrate_vma interfaces, and lib/test_hmm exercises it. migrate.vma is
also populated now for free on fault path. These are aimed at SVM like
migrate-on-fault, and is motivated by Alistair Popple's experiments
with  current interfaces, where the redundant-walk overhead showed up
clearly in perf traces. The unified page table walker also avoids some
limitations in current collecting interfaces that can crash when raced
with unmap.

So this is a building block for migrate-on-fault but delivers immediate
in-tree effects by unifying, refactoring and more robust collecting.

--Mika