Re: [Patch mm-hotfixes v4] mm/page_vma_mapped: fix device-private PMD handling

[Wei Yang]

On Wed, Jun 24, 2026 at 04:57:56PM +0800, Lance Yang wrote:
>
>On Wed, Jun 24, 2026 at 06:53:53AM +0000, Wei Yang wrote:
>>Commit 65edfda6f3f2 ("mm/rmap: extend rmap and migration support
>>device-private entries") introduced the concept of device-private
>>PMD entries, but did not correctly update the rmap walk code to
>>account for them.
>>
>>As a result, when page_vma_mapped_walk() encounters device-private
>>PMD entries, it takes no action other than to acquire the PMD lock
>>and exit.
>>
>>However this is highly problematic for two reasons - firstly,
>>device private entries possess a PFN so check_pmd() needs to be
>>called to ensure an overlapping PFN range.
>>
>>Secondly, and more importantly, if PVMW_MIGRATION is set the
>>caller assumes the returned entry is a migration entry, resulting
>>in memory corruption when the caller tries to interpret the device
>>private entry as such.
>>
>>In addition, commit 146287290023 ("mm/huge_memory: implement
>>device-private THP splitting") allowed device private PMDs to be
>>split like THP mappings, but again did not update this code path.
>>
>>As a result, we might race a PMD split prior to acquiring the PMD
>>lock.
>>
>>This patch addresses all of these issues by invoking check_pmd(),
>>ensuring PMVW_MIGRATION is not set and checks whether a split raced
>>us we do for PMD THP and migration entries.
>>
>>Fixes: 65edfda6f3f2 ("mm/rmap: extend rmap and migration support device-private entries")
>>Cc: <stable@xxxxxxxxxxxxxxx>
>>Signed-off-by: Wei Yang <richard.weiyang@xxxxxxxxx>
>>Suggested-by: David Hildenbrand <david@xxxxxxxxxx>
>
>Shouldn't we add
>
>Suggested-by: Lorenzo Stoakes <ljs@xxxxxxxxxx>
>
>as well?
>
>v4 mostly follows Lorenzo's comments, code bits included. Feels only fair.

Fair enough, added.

>
>>Cc: David Hildenbrand <david@xxxxxxxxxx>
>>Cc: Balbir Singh <balbirs@xxxxxxxxxx>
>>Cc: SeongJae Park <sj@xxxxxxxxxx>
>>Cc: Zi Yan <ziy@xxxxxxxxxx>
>>Cc: Lorenzo Stoakes <ljs@xxxxxxxxxx>
>>Cc: Lance Yang <lance.yang@xxxxxxxxx>
>>
>>---
>>v4:
>>  * refine subject and commit log based on Lorenzo's suggestion
>>  * put pmd device-private entry handling in its own if branch,
>>    suggested by Lorenzo
>>
>>v3:
>>  * remove cleanup part, only fix the issue for device-private entry
>>  * refine user effect description based on Lorenzo's suggestion
>>
>>v2: https://lore.kernel.org/all/20260616063436.20455-1-richard.weiyang@xxxxxxxxx/T/#u
>>  * specify the possible error case of current code and user visible effect
>>  * besides fix, cleanup the pmd entry handling based on David's suggestion
>>
>>v1: https://lore.kernel.org/linux-mm/20260508013728.21285-1-richard.weiyang@xxxxxxxxx/
>>---
>> mm/page_vma_mapped.c | 20 +++++++++++++++-----
>> 1 file changed, 15 insertions(+), 5 deletions(-)
>>
>>diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
>>index 2ccbabfb2cc1..17dff8aab9f9 100644
>>--- a/mm/page_vma_mapped.c
>>+++ b/mm/page_vma_mapped.c
>>@@ -269,14 +269,24 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
>
>
>Hmm ... looks like there may still be a race here ...
>
>Current code picks the branch from the lockless PMD value:
>
>		pmde = pmdp_get_lockless(pvmw->pmd);
>
>		if (pmd_trans_huge(pmde) || pmd_is_migration_entry(pmde)) {
>			pvmw->ptl = pmd_lock(mm, pvmw->pmd);
>			pmde = *pvmw->pmd;
>			if (!pmd_present(pmde)) {
>				softleaf_t entry;
>
>				if (!thp_migration_supported() ||
>				    !(pvmw->flags & PVMW_MIGRATION))
>					return not_found(pvmw);
>				entry = softleaf_from_pmd(pmde);
>
>				if (!softleaf_is_migration(entry) ||
>				    !check_pmd(softleaf_to_pfn(entry), pvmw))
>					return not_found(pvmw);
>				return true;
>			}
>		}
>
>But after taking PTL, the PMD may already be a different non-present PMD
>type:
>
>CPU0: pmde = pmdp_get_lockless();   // sees PMD migration entry
>
>CPU1: remove_migration_ptes(src, dst /* device-private */)
>        ... via rmap_walk(dst) ...
>        page_vma_mapped_walk(&pvmw /* src, PVMW_MIGRATION */)
>          returns with PTL held for the PMD migration entry
>        remove_migration_pmd(new = dst page)
>          installs a device-private PMD
>        next page_vma_mapped_walk()
>          drops PTL via not_found()
>
>CPU0: takes PTL
>      pmde = *pvmw->pmd;            // now device-private PMD
>
>So when PVMW_MIGRATION is not set, current code can return not_found()
>before we even decode the locked PMD as a device-private entry.
>
>Commit 65edfda6f3f2 ("mm/rmap: extend rmap and migration support
>device-private entries") made the
>
>device-private PMD <-> PMD migration
>
>transition possible.
>
>set_pmd_migration_entry() can replace a device-private PMD with a PMD
>migration entry, and remove_migration_pmd() can restore a PMD migration
>entry back to a device-private PMD when the new folio is device-private.
>

Nice catch.

But I think this matters if migration fail and restore the pmd to src folio.
When we successfully migrate to new folio, check_pmd() could catch it and
return not_found(). IIUC.

One more question: assume A unmap a folio, and B migrate the same one. 

If B set_pmd_migration_entry() first, then A won't see this PMD from
page_vma_mapped_walk(), IIUC. Then B failed to migrate, and restore the folio
as this PMD migration entry is there. So A should check the status after
unmap, right? Would it see unstable status?

I am a little lost what is the correct way to do here.

>Maybe decode the locked softleaf entry first, before the migration-only
>checks? Something like this on top:
>
>---8<---
>diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
>index 17dff8aab9f9..97babd408dba 100644
>--- a/mm/page_vma_mapped.c
>+++ b/mm/page_vma_mapped.c
>@@ -249,10 +249,18 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
> 			if (!pmd_present(pmde)) {
> 				softleaf_t entry;
>
>+				entry = softleaf_from_pmd(pmde);
>+				if (softleaf_is_device_private(entry)) {
>+					if (pvmw->flags & PVMW_MIGRATION)
>+						return not_found(pvmw);
>+					if (!check_pmd(softleaf_to_pfn(entry), pvmw))
>+						return not_found(pvmw);
>+					return true;
>+				}
>+

If we have to do this, I am afraid we can put all three cases handling
here...

Not necessary to put pmd_is_device_private_entry() handling in two places.

> 				if (!thp_migration_supported() ||
> 				    !(pvmw->flags & PVMW_MIGRATION))
> 					return not_found(pvmw);
>-				entry = softleaf_from_pmd(pmde);
>
> 				if (!softleaf_is_migration(entry) ||
> 				    !check_pmd(softleaf_to_pfn(entry), pvmw))
>@@ -266,7 +274,10 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
> 					return not_found(pvmw);
> 				return true;
> 			}
>-			/* THP pmd was split under us: handle on pte level */
>+			/*
>+			 * THP pmd was split under us, or device-private PMD
>+			 * changed under us: handle on pte level.
>+			 */
> 			spin_unlock(pvmw->ptl);
> 			pvmw->ptl = NULL;
> 		} else if (pmd_is_device_private_entry(pmde)) {
>--
>
>Anyway, that stuff is getting kinda messy now. Feels like it really needs
>a cleanup on top before it bites us again :)

Agree.

I haven't imagined this would be more complicated than I thought :-)

>Cheers, Lance

-- 
Wei Yang
Help you, Help me