Re: [PATCH v14 10/12] mm: enable device page migration from HMM pagewalk

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] mm: unbalanced mmu_notifier in hmm_range_fault during migration
--

commit 3e05399d761177583ed386d43f42e228c070f447
Author: Mika Penttilä <mpenttil@redhat.com>

mm: enable device page migration from HMM pagewalk

This patch integrates HMM pagewalk machinery into the first phase of device
page migration by enabling migration in hmm_range_fault(). It updates
migrate_vma_setup() to use the HMM pagewalk path and removes the older
migrate_vma_collect() paths. Two new flags, MIGRATE_VMA_FAULT and
MIGRATE_VMA_WRITE, are introduced for faulting missing pages and requesting
write access.

> diff --git a/mm/migrate_device.c b/mm/migrate_device.c
> index 43e571e82ccad..1e1a592e12672 100644
> --- a/mm/migrate_device.c
> +++ b/mm/migrate_device.c
[ ... ]
> @@ -739,10 +241,20 @@ static void migrate_vma_unmap(struct migrate_vma *migrate)
>   */
>  int migrate_vma_setup(struct migrate_vma *args)
>  {
> +	int ret;
>  	long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
> +	struct hmm_range range = {
> +		.notifier = NULL,
> +		.hmm_pfns = args->src,
> +		.dev_private_owner = args->pgmap_owner,

[Severity: High]
Does this code cause a permanent MMU notifier imbalance when args->pgmap_owner
is NULL?

If a caller invokes migrate_vma_setup() with args->pgmap_owner set to NULL
(which is standard for migrations not involving device-private memory, such as
PowerPC KVM secure guest page-in), range.dev_private_owner is also NULL.

In mm/hmm.c:hmm_vma_capture_migrate_range(), the code checks:

	if (!hmm_vma_walk->mmu_range.owner) {
		mmu_notifier_range_init_owner(&hmm_vma_walk->mmu_range, MMU_NOTIFY_MIGRATE, 0,
					      walk->vma->vm_mm, start, end,
					      range->dev_private_owner);
		mmu_notifier_invalidate_range_start(&hmm_vma_walk->mmu_range);
	}

Because mmu_range is zero-initialized, this condition is true. It calls
mmu_notifier_range_init_owner(), which assigns the NULL dev_private_owner
to hmm_vma_walk->mmu_range.owner. mmu_notifier_invalidate_range_start() is
then called.

Additionally, if the range spans multiple VMAs, invalidate_range_start is
called multiple times because the sentinel remains NULL.

Later, in mm/hmm.c:hmm_vma_post_range_fault():

	if (hmm_select_migrate(range) &&
	    hmm_vma_walk->mmu_range.owner) {
		mmu_notifier_invalidate_range_end(&hmm_vma_walk->mmu_range);
	}

Because mmu_range.owner remains NULL, this condition evaluates to false,
completely skipping the required mmu_notifier_invalidate_range_end() call.

A skipped invalidate_range_end() leaves the invalidation count permanently
elevated, causing any subsequent operation on that memory range requiring
mmu_interval_read_retry() or VMA teardown to block indefinitely.

> +		.migrate = args,
> +		.default_flags = HMM_PFN_REQ_MIGRATE
> +	};
>  
>  	args->start &= PAGE_MASK;
>  	args->end &= PAGE_MASK;
[ ... ]

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260922053421.4092027-1-mpenttil@redhat.com?part=10