Re: [PATCH] mm/page_vma_mapped: revalidate and do proper check before return device-private pmd

[Lorenzo Stoakes]

On Mon, Jun 22, 2026 at 11:45:18PM +0000, Wei Yang wrote:
> On Mon, Jun 22, 2026 at 05:11:02PM +0100, Lorenzo Stoakes wrote:
> >On Mon, Jun 22, 2026 at 02:21:02PM +0000, Wei Yang wrote:
> >> On Mon, Jun 22, 2026 at 02:46:40PM +0100, Lorenzo Stoakes wrote:
> >> >+cc Lance, linux-kernel
> >> >
> >> >Your subject line is 83 characters long and is way too detailed how about 'fix
> >> >device-private PMD handling'?
> >> >
> >>
> >> Got it.
> >>
> >> >You forgot to include linux-kernel@xxxxxxxxxxxxxxx on the mail, lore seems to be
> >> >a bit broken atm but in general it's helpful to include that.
> >>
> >> Got it.
> >>
> >> So usually we send a patch to both linux-mm and linux-kernel? If so, I
> >> remember is later actions.
> >
> >Yeah it's better for dealing with kvack going wrong etc. :)
> >
> >>
> >> >
> >> >Also is useful to make this [PATCH mm-hotfixes] to make it really clear it's
> >> >intended as a hotfix.
> >> >
> >>
> >> Got it.
> >>
> >> >Some commit msg language nits:
> >> >
> >> >On Mon, Jun 22, 2026 at 01:06:51PM +0000, Wei Yang wrote:
> >> >> For pmd_trans_huge() and pmd_is_migration_entry(), we does following
> >> >> before return the pmd entry:
> >> >
> >> >Sounds better as:
> >> >
> >> >	For PMD entries that satisfy pmd_trans_huge() or pmd_is_migration_entry(), we
> >> >	perform the following actions:
> >> >
> >>
> >> Sure.
> >>
> >> >>
> >> >>   * re-validate pmd entry after PTL
> >> >>   * check PVMW_MIGRATION
> >> >>   * check_pmd()
> >> >>   * handle on pte level if split under us
> >> >>
> >> >> But for device-private pmd, we just return after pmd_lock().
> >> >
> >> >->
> >> >
> >> >	However, for device-private PMD entries, we simply acquire the PMD lock
> >> >	and return.
> >> >
> >>
> >> Sure.
> >>
> >> >Also can you please give some justification here as to why all this also applies
> >> >to device-private PMD? Right now it sounds hand wavey.
> >> >
> >>
> >> I thought below paragraph explain it. Not sure what justification is preferred.
> >
> >Something about device private PMDs splitting the same way THP ones do, in the
> >pmd_is_device_private_entry() branch of __split_huge_pmd_locked().
> >
>
> Hi, Lorenzo
>
> Thanks for your detailed suggestions.
>
> I tried to add the justification here, and the following is the commit log
> after consolidate your suggestions.
>
>     For PMD entries that satisfy pmd_trans_huge() or
>     pmd_is_migration_entry(), we perform the following actions:
>
>       * re-validate pmd entry after PTL
>       * check PVMW_MIGRATION
>       * check_pmd()
>       * handle on pte level if split under us
>
>     However, for device-private PMD entries, we simply acquire the PMD lock
>     and return. This is not enough, as __split_huge_pmd_locked() would split
>     a pmd device-private PMD under us just as it does for THP PMD.
>
>     This is particularly problematic when PVMW_MIGRATION is set (meaning a
>     migration entry is sought), as it causes a device-private PMD entry to
>     be returned with a different data layout, causing memory corruption.
>
> Just feel this is not that smooth. Would you mind taking another look to see
> if I get your point correctly?

Honestly I'd just drop the whole pmd_trans_huge()/pmd_is_migration_entry() bit
and say:

	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.

>
> --
> Wei Yang
> Help you, Help me

Cheers, Lorenzo