Re: [PATCH] mm/hmm: account for subpage offsets in device-private PMDs
From: Mika Penttilä
Date: Sun Nov 22 2026 - 00:55:51 EST
On 11/22/26 02:00, Andrew Morton wrote:
> On Sat, 31 Oct 2026 15:07:58 +1100 Aveline Noir <jm5905938@xxxxxxxxx> wrote:
>
>> When a query starts inside a device-private PMD owned by
>> range->dev_private_owner, hmm_vma_handle_absent_pmd() still fills the PFN
>> array from the folio's first page. The returned PFNs therefore refer to
>> earlier pages rather than the requested subrange.
>>
>> Add the query's PMD-relative page offset to the initial PFN, as the
>> present PMD path already does.
>>
>> Fixes: 10b9feee2d0d ("mm/hmm: populate PFNs from PMD swap entry")
>> Signed-off-by: Aveline Noir <jm5905938@xxxxxxxxx>
>>
>> ---
>> Found during a synthetic test_hmm check of device-private THP handling.
>> Tested PMD-aligned and non-aligned subranges, PMD boundaries, split
>> PTEs and fake-device data accesses with test_hmm in QEMU TCG; all
>> executed tests passed with the fix. No real GPU testing was performed.
> Thanks. It appears that the changelog somewhat understates the
> seriousness of this so I came up with a more assertive one, below.
> Please check that.
>
> I fed this through LLM-driven QEMU-based runtime testing and it looks
> like it pretty much replicated your testing:
>
> : I tested "mm/hmm: account for subpage offsets in device-private
> : PMDs" on matched x86-64 guest kernels. An unprivileged process
> : (UID 65534) migrated an anonymous 2 MiB THP to the test_hmm
> : device. A second open of the device read pages inside the PMD
> : while the first remained open. A snapshot confirmed a
> : device-private PMD, and temporary kernel log messages confirmed
> : that the reads reached hmm_vma_handle_absent_pmd().
> :
> : Without the patch, requests for pages 1, 7, 255 and 511 returned
> : page 0. A request for pages 31-33 returned pages 0-2. With the
> : patch, all seven pages returned the requested contents. Page 0
> : read correctly on both kernels. Thus an unprivileged device user
> : can receive data from the wrong pages.
> :
> : The fake device needed an identical temporary backing-page
> : correction on both kernels to make this read path work. That
> : change was removed after testing. This exercises the HMM path
> : through test_hmm; I did not test GPU hardware.
>
> So it accidentally found a bug in lib/test_hmm.c on the way. "The QEMU
> run exposed a real test_hmm bug: a second open of /dev/hmm_dmirror0 can
> put a device-private page into its page table, then dmirror_do_read()
> tries to copy from that non-CPU-accessible page and oopses". I'll prep
> a fix for that separately.
I noticed that also and will make changes around it in my series.
>
>
>
> I'll queue the below in mm.git's mm-hotfixes queue and shall await
> maintainer input (please).
>
>
> From: Aveline Noir <jm5905938@xxxxxxxxx>
> Subject: mm/hmm: account for subpage offsets in device-private PMDs
> Date: Sat, 31 Oct 2026 15:07:58 +1100
>
> When a query starts inside a device-private PMD owned by
> range->dev_private_owner, hmm_vma_handle_absent_pmd() still fills the PFN
> array from the folio's first page. The returned PFNs therefore refer to
> earlier pages rather than the requested subrange.
>
> Nouveau's GPU fault path can query one page of a device-private PMD at a
> time. Xe's SVM fault, prefetch and userptr mapping paths can also query
> subranges of their own device-private PMDs. An unprivileged process with
> access to the GPU render node can reach these paths on hardware supporting
> device-private THP migration. Xe can read the wrong application data or
> silently overwrite earlier pages of the same process's device-private
> folio.
>
> Nouveau aligns a one-page GPU fault to the PMD head by subtracting the
> fault offset from the returned page. The faulty HMM PFN already refers to
> the head, so this subtraction can select a page before the folio. If the
> preceding page belongs to another allocation, GPU access can expose or
> overwrite unrelated userspace data; an invalid page can also fault the
> kernel during its address conversion. This worst case follows from the
> source, but has not been demonstrated on a real GPU.
>
> Add the query's PMD-relative page offset to the initial PFN, as the
> present PMD path already does.
>
> The device-private PMD migration machinery first shipped in 6.19.
> Backport this fix to the maintained 7.3.y stable series; the HMM path
> exists in 6.18.y but cannot be reached through an in-tree PMD-producing
> device-private migration path there.
>
> Link: https://patch.msgid.link/179342234355.4.16705453209886955762@xxxxxxxxx
> Fixes: 10b9feee2d0d ("mm/hmm: populate PFNs from PMD swap entry")
> Signed-off-by: Aveline Noir <jm5905938@xxxxxxxxx>
> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Jason Gunthorpe <jgg@xxxxxxxx>
> Cc: Leon Romanovsky <leon@xxxxxxxxxx>
> Cc: Francois Dugast <francois.dugast@xxxxxxxxx>
> Cc: Lorenzo Stoakes <ljs@xxxxxxxxxx>
> Cc: Matthew Brost <matthew.brost@xxxxxxxxx>
> Cc: Patchwork <patchwork@xxxxxxxxxxxxxxxxxxxxxx>
> Cc: SJ Park <sj@xxxxxxxxxx>
> Cc: Vlastimil Babka <vbabka@xxxxxxxxxx>
> Cc: David Airlie <airlied@xxxxxxxxx>
> Cc: Alistair Popple <apopple@xxxxxxxxxx>
> Cc: Balbir Singh <balbirs@xxxxxxxxxx>
> Cc: Christian König <christian.koenig@xxxxxxx>
> Cc: Mika Penttilä <mpenttil@xxxxxxxxxx>
> Cc: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
> Cc: Zi Yan <ziy@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx>
>
> mm/hmm.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> --- a/mm/hmm.c~mm-hmm-account-for-subpage-offsets-in-device-private-pmds
> +++ a/mm/hmm.c
> @@ -355,7 +355,8 @@ static int hmm_vma_handle_absent_pmd(str
> range->dev_private_owner) {
> unsigned long cpu_flags = HMM_PFN_VALID |
> hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
> - unsigned long pfn = softleaf_to_pfn(entry);
> + unsigned long pfn = softleaf_to_pfn(entry) +
> + ((start & ~PMD_MASK) >> PAGE_SHIFT);
> unsigned long i;
>
> if (softleaf_is_device_private_write(entry))
> _
For this change:
Reviewed-by: Mika Penttilä <mpenttil@xxxxxxxxxx>