Re: [PATCH] lib/test_hmm: fix alloc tag leak when splitting device-private folios
From: Nilay Vaish
Date: Thu Nov 19 2026 - 18:29:06 EST
Zi Yan, thanks for the review. I'm about to send a v2 that only
replaces the pfn_to_page(page_to_pfn())
round trips with plain struct page arithmetic, as you suggested
Zenghui Yu: Thanks for testing the patch! v2 has no functional
changes. If you get a chance to run it, would you mind
replying with a Tested-by tag?
Thanks,
Nilay
On Thu, Nov 19, 2026 at 7:38 AM Zi Yan <ziy@xxxxxxxxxx> wrote:
>
> On 18 Nov 2026, at 18:24, Zi Yan wrote:
>
> > On 18 Nov 2026, at 18:00, Nilay Vaish wrote:
> >
> >> For device-private memory, test_hmm backs each fake device folio with
> >> real system RAM of the same order, allocated with folio_alloc() in
> >> dmirror_devmem_alloc_page().
> >>
> >> When the core MM splits a large device-private folio, the driver's
> >> ->folio_split() callback, dmirror_devmem_folio_split(), has to split
> >> the backing memory as well. It does that by hand: clear_compound_head()
> >> on the tail pages, folio_reset_order() on the head page, and
> >> folio_set_count() on each of them. dmirror_devmem_free() then frees
> >> every backing page as an order-0 page.
> >>
> >> This open-coded split does not split the allocator's metadata. The
> >> allocation tag and the page_owner record still describe one high-order
> >> allocation, and PG_head stays set on the head page. When the pages are
> >> freed one by one, the tag is debited for only one page when the head is
> >> freed, and the tail pages carry no tag at all. With
> >> CONFIG_MEM_ALLOC_PROFILING, running the HMM selftests and unloading the
> >> module reports the leak:
> >>
> >> lib/test_hmm.c:808 module test_hmm func:dmirror_devmem_alloc_page has 16744448 allocated at module unload
> >>
> >> and with CONFIG_MEM_ALLOC_PROFILING_DEBUG freeing a tail page triggers:
> >>
> >> alloc_tag was not set
> >> WARNING: ./include/linux/alloc_tag.h:164 at ___free_pages+0x2a0/0x2d0
> >>
> >> The backing memory is a private buffer of the test driver and nothing
> >> requires it to be a folio. Allocate it with alloc_pages() instead and
> >> split it right away with split_page(), which splits the allocation tag,
> >> page_owner and memcg state and is already exported to modules. It is
> >> still allocated as one high-order block rather than as individual
> >> order-0 pages because the driver finds the backing page of the Nth page
> >> of a device folio by offsetting from the first backing page, so the
> >> backing pages have to be physically contiguous. Every backing page is
> >> then an independent order-0 page for its whole lifetime, so
> >> dmirror_devmem_folio_split() only has to point the new device folio at
> >> its backing page, and dmirror_devmem_free() frees the backing pages one
> >> at a time.
> >>
> >> As the backing pages no longer form a folio, take the order from the
> >> device folio rather than from the backing page, and index and copy the
> >> backing pages individually instead of using folio_page() and
> >> folio_copy().
> >>
> >> With this patch the device-private HMM selftests pass and the module
> >> unloads without the messages above.
> >>
> >> Fixes: 56ef39899643 ("mm/memremap: add driver callback support for folio splitting")
> >> Reported-by: Zenghui Yu <zenghui.yu@xxxxxxxxx>
> >> Closes: https://lore.kernel.org/linux-mm/be9a8bae-9223-4966-bba9-2cbe39c8f4de@xxxxxxxxx/
> >> Suggested-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
> >> Link: https://lore.kernel.org/linux-mm/8ef6e92d-4cb3-444e-857e-fdbf88c035ea@xxxxxxxxxx/
> >> Signed-off-by: Nilay Vaish <nilayvaish@xxxxxxxxxx>
> >> ---
> >> lib/test_hmm.c | 41 +++++++++++++++++------------------------
> >> 1 file changed, 17 insertions(+), 24 deletions(-)
> >
> >
> >>
> >> diff --git a/lib/test_hmm.c b/lib/test_hmm.c
> >> index b409c42bfe6f..371ead009f8a 100644
> >> --- a/lib/test_hmm.c
> >> +++ b/lib/test_hmm.c
> >> @@ -201,6 +201,7 @@ static void dmirror_device_evict_chunk(struct dmirror_chunk *chunk)
> >> migrate_device_range(src_pfns, start_pfn, npages);
> >> for (i = 0; i < npages; i++) {
> >> struct page *dpage, *spage;
> >> + unsigned int j;
> >>
> >> spage = migrate_pfn_to_page(src_pfns[i]);
> >> if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE))
> >> @@ -227,7 +228,9 @@ static void dmirror_device_evict_chunk(struct dmirror_chunk *chunk)
> >> dst_pfns[i] |= MIGRATE_PFN_WRITE;
> >> if (order)
> >> dst_pfns[i] |= MIGRATE_PFN_COMPOUND;
> >> - folio_copy(page_folio(dpage), page_folio(spage));
> >> + for (j = 0; j < (1U << order); j++)
> >> + copy_highpage(pfn_to_page(page_to_pfn(dpage) + j),
> >> + pfn_to_page(page_to_pfn(spage) + j));
> >
> > struct page should be contiguous if their PFNs are contiguous.
> > There is no need to do this conversion, right?
> >
> > In addition, dpage, based on the comment before BACKING_PAGE,
> > is device memory, so it should also be allocated using alloc_pages()
> > like rpage. Am I getting it wrong?
>
> I misread the comment, dpage is destination page. It can be ZONE_DEVICE
> page or system memory depending on functions.
>
>
> Best Regards,
> Yan, Zi