Message-ID: <0C16A79F-5A7B-4358-9806-7F78E7EA8EE6@nvidia.com>
Date: Thu, 22 Jan 2026 22:09:41 -0500
From: Zi Yan <ziy@...dia.com>
To: Alistair Popple <apopple@...dia.com>
Cc: Jordan Niethe <jniethe@...dia.com>, <linux-mm@...ck.org>,
 <balbirs@...dia.com>, <matthew.brost@...el.com>, <akpm@...ux-foundation.org>,
 <linux-kernel@...r.kernel.org>, <dri-devel@...ts.freedesktop.org>,
 <david@...hat.com>, <lorenzo.stoakes@...cle.com>, <lyude@...hat.com>,
 <dakr@...nel.org>, <airlied@...il.com>, <simona@...ll.ch>,
 <rcampbell@...dia.com>, <mpenttil@...hat.com>, <jgg@...dia.com>,
 <willy@...radead.org>, <linuxppc-dev@...ts.ozlabs.org>,
 <intel-xe@...ts.freedesktop.org>, <jgg@...pe.ca>, <Felix.Kuehling@....com>
Subject: Re: [PATCH v2 11/11] mm: Remove device private pages from the
 physical address space

On 22 Jan 2026, at 22:06, Zi Yan wrote:

> On 22 Jan 2026, at 21:02, Alistair Popple wrote:
>
>> On 2026-01-21 at 10:06 +1100, Zi Yan <ziy@...dia.com> wrote...
>>> On 20 Jan 2026, at 18:02, Jordan Niethe wrote:
>>>
>>>> Hi,
>>>>
>>>> On 21/1/26 09:53, Zi Yan wrote:
>>>>> On 20 Jan 2026, at 17:33, Jordan Niethe wrote:
>>>>>
>>>>>> On 14/1/26 07:04, Zi Yan wrote:
>>>>>>> On 7 Jan 2026, at 4:18, Jordan Niethe wrote:
>>>>>>>
>>>>>>>> Currently when creating these device private struct pages, the first
>>>>>>>> step is to use request_free_mem_region() to get a range of physical
>>>>>>>> address space large enough to represent the devices memory. This
>>>>>>>> allocated physical address range is then remapped as device private
>>>>>>>> memory using memremap_pages().
>>>>>>>>
>>>>>>>> Needing allocation of physical address space has some problems:
>>>>>>>>
>>>>>>>>     1) There may be insufficient physical address space to represent the
>>>>>>>>        device memory. KASLR reducing the physical address space and VM
>>>>>>>>        configurations with limited physical address space increase the
>>>>>>>>        likelihood of hitting this especially as device memory increases. This
>>>>>>>>        has been observed to prevent device private from being initialized.
>>>>>>>>
>>>>>>>>     2) Attempting to add the device private pages to the linear map at
>>>>>>>>        addresses beyond the actual physical memory causes issues on
>>>>>>>>        architectures like aarch64 meaning the feature does not work there.
>>>>>>>>
>>>>>>>> Instead of using the physical address space, introduce a device private
>>>>>>>> address space and allocate devices regions from there to represent the
>>>>>>>> device private pages.
>>>>>>>>
>>>>>>>> Introduce a new interface memremap_device_private_pagemap() that
>>>>>>>> allocates a requested amount of device private address space and creates
>>>>>>>> the necessary device private pages.
>>>>>>>>
>>>>>>>> To support this new interface, struct dev_pagemap needs some changes:
>>>>>>>>
>>>>>>>>     - Add a new dev_pagemap::nr_pages field as an input parameter.
>>>>>>>>     - Add a new dev_pagemap::pages array to store the device
>>>>>>>>       private pages.
>>>>>>>>
>>>>>>>> When using memremap_device_private_pagemap(), rather then passing in
>>>>>>>> dev_pagemap::ranges[dev_pagemap::nr_ranges] of physical address space to
>>>>>>>> be remapped, dev_pagemap::nr_ranges will always be 1, and the device
>>>>>>>> private range that is reserved is returned in dev_pagemap::range.
>>>>>>>>
>>>>>>>> Forbid calling memremap_pages() with dev_pagemap::ranges::type =
>>>>>>>> MEMORY_DEVICE_PRIVATE.
>>>>>>>>
>>>>>>>> Represent this device private address space using a new
>>>>>>>> device_private_pgmap_tree maple tree. This tree maps a given device
>>>>>>>> private address to a struct dev_pagemap, where a specific device private
>>>>>>>> page may then be looked up in that dev_pagemap::pages array.
>>>>>>>>
>>>>>>>> Device private address space can be reclaimed and the assoicated device
>>>>>>>> private pages freed using the corresponding new
>>>>>>>> memunmap_device_private_pagemap() interface.
>>>>>>>>
>>>>>>>> Because the device private pages now live outside the physical address
>>>>>>>> space, they no longer have a normal PFN. This means that page_to_pfn(),
>>>>>>>> et al. are no longer meaningful.
>>>>>>>>
>>>>>>>> Introduce helpers:
>>>>>>>>
>>>>>>>>     - device_private_page_to_offset()
>>>>>>>>     - device_private_folio_to_offset()
>>>>>>>>
>>>>>>>> to take a given device private page / folio and return its offset within
>>>>>>>> the device private address space.
>>>>>>>>
>>>>>>>> Update the places where we previously converted a device private page to
>>>>>>>> a PFN to use these new helpers. When we encounter a device private
>>>>>>>> offset, instead of looking up its page within the pagemap use
>>>>>>>> device_private_offset_to_page() instead.
>>>>>>>>
>>>>>>>> Update the existing users:
>>>>>>>>
>>>>>>>>    - lib/test_hmm.c
>>>>>>>>    - ppc ultravisor
>>>>>>>>    - drm/amd/amdkfd
>>>>>>>>    - gpu/drm/xe
>>>>>>>>    - gpu/drm/nouveau
>>>>>>>>
>>>>>>>> to use the new memremap_device_private_pagemap() interface.
>>>>>>>>
>>>>>>>> Signed-off-by: Jordan Niethe <jniethe@...dia.com>
>>>>>>>> Signed-off-by: Alistair Popple <apopple@...dia.com>
>>>>>>>>
>>>>>>>> ---
>>>>>>>>
>>>>>>>> NOTE: The updates to the existing drivers have only been compile tested.
>>>>>>>> I'll need some help in testing these drivers.
>>>>>>>>
>>>>>>>> v1:
>>>>>>>> - Include NUMA node paramater for memremap_device_private_pagemap()
>>>>>>>> - Add devm_memremap_device_private_pagemap() and friends
>>>>>>>> - Update existing users of memremap_pages():
>>>>>>>>       - ppc ultravisor
>>>>>>>>       - drm/amd/amdkfd
>>>>>>>>       - gpu/drm/xe
>>>>>>>>       - gpu/drm/nouveau
>>>>>>>> - Update for HMM huge page support
>>>>>>>> - Guard device_private_offset_to_page and friends with CONFIG_ZONE_DEVICE
>>>>>>>>
>>>>>>>> v2:
>>>>>>>> - Make sure last member of struct dev_pagemap remains DECLARE_FLEX_ARRAY(struct range, ranges);
>>>>>>>> ---
>>>>>>>>    Documentation/mm/hmm.rst                 |  11 +-
>>>>>>>>    arch/powerpc/kvm/book3s_hv_uvmem.c       |  41 ++---
>>>>>>>>    drivers/gpu/drm/amd/amdkfd/kfd_migrate.c |  23 +--
>>>>>>>>    drivers/gpu/drm/nouveau/nouveau_dmem.c   |  35 ++--
>>>>>>>>    drivers/gpu/drm/xe/xe_svm.c              |  28 +---
>>>>>>>>    include/linux/hmm.h                      |   3 +
>>>>>>>>    include/linux/leafops.h                  |  16 +-
>>>>>>>>    include/linux/memremap.h                 |  64 +++++++-
>>>>>>>>    include/linux/migrate.h                  |   6 +-
>>>>>>>>    include/linux/mm.h                       |   2 +
>>>>>>>>    include/linux/rmap.h                     |   5 +-
>>>>>>>>    include/linux/swapops.h                  |  10 +-
>>>>>>>>    lib/test_hmm.c                           |  69 ++++----
>>>>>>>>    mm/debug.c                               |   9 +-
>>>>>>>>    mm/memremap.c                            | 193 ++++++++++++++++++-----
>>>>>>>>    mm/mm_init.c                             |   8 +-
>>>>>>>>    mm/page_vma_mapped.c                     |  19 ++-
>>>>>>>>    mm/rmap.c                                |  43 +++--
>>>>>>>>    mm/util.c                                |   5 +-
>>>>>>>>    19 files changed, 391 insertions(+), 199 deletions(-)
>>>>>>>>
>>>>>>> <snip>
>>>>>>>
>>>>>>>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>>>>>>>> index e65329e1969f..b36599ab41ba 100644
>>>>>>>> --- a/include/linux/mm.h
>>>>>>>> +++ b/include/linux/mm.h
>>>>>>>> @@ -2038,6 +2038,8 @@ static inline unsigned long memdesc_section(memdesc_flags_t mdf)
>>>>>>>>     */
>>>>>>>>    static inline unsigned long folio_pfn(const struct folio *folio)
>>>>>>>>    {
>>>>>>>> +	VM_BUG_ON(folio_is_device_private(folio));
>>>>>>>
>>>>>>> Please use VM_WARN_ON instead.
>>>>>>
>>>>>> ack.
>>>>>>
>>>>>>>
>>>>>>>> +
>>>>>>>>    	return page_to_pfn(&folio->page);
>>>>>>>>    }
>>>>>>>>
>>>>>>>> diff --git a/include/linux/rmap.h b/include/linux/rmap.h
>>>>>>>> index 57c63b6a8f65..c1561a92864f 100644
>>>>>>>> --- a/include/linux/rmap.h
>>>>>>>> +++ b/include/linux/rmap.h
>>>>>>>> @@ -951,7 +951,7 @@ static inline unsigned long page_vma_walk_pfn(unsigned long pfn)
>>>>>>>>    static inline unsigned long folio_page_vma_walk_pfn(const struct folio *folio)
>>>>>>>>    {
>>>>>>>>    	if (folio_is_device_private(folio))
>>>>>>>> -		return page_vma_walk_pfn(folio_pfn(folio)) |
>>>>>>>> +		return page_vma_walk_pfn(device_private_folio_to_offset(folio)) |
>>>>>>>>    		       PVMW_PFN_DEVICE_PRIVATE;
>>>>>>>>
>>>>>>>>    	return page_vma_walk_pfn(folio_pfn(folio));
>>>>>>>> @@ -959,6 +959,9 @@ static inline unsigned long folio_page_vma_walk_pfn(const struct folio *folio)
>>>>>>>>
>>>>>>>>    static inline struct page *page_vma_walk_pfn_to_page(unsigned long pvmw_pfn)
>>>>>>>>    {
>>>>>>>> +	if (pvmw_pfn & PVMW_PFN_DEVICE_PRIVATE)
>>>>>>>> +		return device_private_offset_to_page(pvmw_pfn >> PVMW_PFN_SHIFT);
>>>>>>>> +
>>>>>>>>    	return pfn_to_page(pvmw_pfn >> PVMW_PFN_SHIFT);
>>>>>>>>    }
>>>>>>>
>>>>>>> <snip>
>>>>>>>
>>>>>>>> diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
>>>>>>>> index 96c525785d78..141fe5abd33f 100644
>>>>>>>> --- a/mm/page_vma_mapped.c
>>>>>>>> +++ b/mm/page_vma_mapped.c
>>>>>>>> @@ -107,6 +107,7 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
>>>>>>>>    static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
>>>>>>>>    {
>>>>>>>>    	unsigned long pfn;
>>>>>>>> +	bool device_private = false;
>>>>>>>>    	pte_t ptent = ptep_get(pvmw->pte);
>>>>>>>>
>>>>>>>>    	if (pvmw->flags & PVMW_MIGRATION) {
>>>>>>>> @@ -115,6 +116,9 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
>>>>>>>>    		if (!softleaf_is_migration(entry))
>>>>>>>>    			return false;
>>>>>>>>
>>>>>>>> +		if (softleaf_is_migration_device_private(entry))
>>>>>>>> +			device_private = true;
>>>>>>>> +
>>>>>>>>    		pfn = softleaf_to_pfn(entry);
>>>>>>>>    	} else if (pte_present(ptent)) {
>>>>>>>>    		pfn = pte_pfn(ptent);
>>>>>>>> @@ -127,8 +131,14 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
>>>>>>>>    			return false;
>>>>>>>>
>>>>>>>>    		pfn = softleaf_to_pfn(entry);
>>>>>>>> +
>>>>>>>> +		if (softleaf_is_device_private(entry))
>>>>>>>> +			device_private = true;
>>>>>>>>    	}
>>>>>>>>
>>>>>>>> +	if ((device_private) ^ !!(pvmw->pfn & PVMW_PFN_DEVICE_PRIVATE))
>>>>>>>> +		return false;
>>>>>>>> +
>>>>>>>>    	if ((pfn + pte_nr - 1) < (pvmw->pfn >> PVMW_PFN_SHIFT))
>>>>>>>>    		return false;
>>>>>>>>    	if (pfn > ((pvmw->pfn >> PVMW_PFN_SHIFT) + pvmw->nr_pages - 1))
>>>>>>>> @@ -137,8 +147,11 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
>>>>>>>>    }
>>>>>>>>
>>>>>>>>    /* Returns true if the two ranges overlap.  Careful to not overflow. */
>>>>>>>> -static bool check_pmd(unsigned long pfn, struct page_vma_mapped_walk *pvmw)
>>>>>>>> +static bool check_pmd(unsigned long pfn, bool device_private, struct page_vma_mapped_walk *pvmw)
>>>>>>>>    {
>>>>>>>> +	if ((device_private) ^ !!(pvmw->pfn & PVMW_PFN_DEVICE_PRIVATE))
>>>>>>>> +		return false;
>>>>>>>> +
>>>>>>>>    	if ((pfn + HPAGE_PMD_NR - 1) < (pvmw->pfn >> PVMW_PFN_SHIFT))
>>>>>>>>    		return false;
>>>>>>>>    	if (pfn > (pvmw->pfn >> PVMW_PFN_SHIFT) + pvmw->nr_pages - 1)
>>>>>>>> @@ -255,6 +268,8 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
>>>>>>>>
>>>>>>>>    				if (!softleaf_is_migration(entry) ||
>>>>>>>>    				    !check_pmd(softleaf_to_pfn(entry),
>>>>>>>> +					       softleaf_is_device_private(entry) ||
>>>>>>>> +					       softleaf_is_migration_device_private(entry),
>>>>>>>>    					       pvmw))
>>>>>>>>    					return not_found(pvmw);
>>>>>>>>    				return true;
>>>>>>>> @@ -262,7 +277,7 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
>>>>>>>>    			if (likely(pmd_trans_huge(pmde))) {
>>>>>>>>    				if (pvmw->flags & PVMW_MIGRATION)
>>>>>>>>    					return not_found(pvmw);
>>>>>>>> -				if (!check_pmd(pmd_pfn(pmde), pvmw))
>>>>>>>> +				if (!check_pmd(pmd_pfn(pmde), false, pvmw))
>>>>>>>>    					return not_found(pvmw);
>>>>>>>>    				return true;
>>>>>>>>    			}
>>>>>>>
>>>>>>> It seems to me that you can add a new flag like “bool is_device_private” to
>>>>>>> indicate whether pfn is a device private index instead of pfn without
>>>>>>> manipulating pvmw->pfn itself.
>>>>>>
>>>>>> We could do it like that, however my concern with using a new param was that
>>>>>> storing this info seperately might make it easier to misuse a device private
>>>>>> index as a regular pfn.
>>>>>>
>>>>>> It seemed like it could be easy to overlook both when creating the pvmw and
>>>>>> then when accessing the pfn.
>>>>>
>>>>> That is why I asked for a helper function like page_vma_walk_pfn(pvmw) to
>>>>> return the converted pfn instead of pvmw->pfn directly. You can add a comment
>>>>> to ask people to use helper function and even mark pvmw->pfn /* do not use
>>>>> directly */.
>>>>
>>>> Yeah I agree that is a good idea.
>>>>
>>>>>
>>>>> In addition, your patch manipulates pfn by left shifting it by 1. Are you sure
>>>>> there is no weird arch having pfns with bit 63 being 1? Your change could
>>>>> break it, right?
>>>>
>>>> Currently for migrate pfns we left shift by pfns by MIGRATE_PFN_SHIFT (6), so I
>>>> thought doing something similiar here should be safe.
>>>
>>> Yeah, but that limits to archs supporting HMM. page_vma_mapped_walk is used
>>> by almost every arch, so it has a broader impact.
>>
>> We need to be a bit careful by what we mean when we say "HMM" in the kernel.
>>
>> Specifically MIGRATE_PFN_SHIFT is used with migrate_vma/migrate_device, which
>> is the migration half of "HMM" which does depend on CONFIG_DEVICE_MIGRATION or
>> really just CONFIG_ZONE_DEVICE making it somewhat arch specific.
>>
>> However hmm_range_fault() does something similar - see the definition of
>> hmm_pfn_flags - it actually steals the top 11 bits of a pfn for flags, and it is
>> not architecture specific. It only depends on CONFIG_MMU.
>
> Oh, that is hacky. But are HMM PFNs with any flag exposed to code outside HMM?
> Currently, device private needs to reserve PFNs for struct page, so I assume
> only the reserved PFNs are seen by outsiders. Otherwise, when outsiders see
> a HMM PFN with a flag, pfn_to_page() on such a PFN will read non exist
> struct page, right?
>
> For this page_vma_mapped_walk code, it is manipulating PFNs used by everyone,
> not just HMM, and can potentially (might be very rare) alter their values
> after shifts.


> And if an HMM PFN with HMM_PFN_VALID is processed by the code,
> the HMM PFN will lose HMM_PFN_VALID bit. So I guess HMM PFN is not showing
> outside HMM code.

Oops, this code is removing PFN reservation mechanism, so please ignore
the above two sentences.

>
>>
>> Now I'm not saying this implies it actually works on all architectures as I
>> agree the page_vma_mapped_walk code is used much more widely. Rather I'm just
>> pointing out if there are issues with some architectures using high PFN bits
>> then we likely have a problem here too :-)
>
>
> Best Regards,
> Yan, Zi


Best Regards,
Yan, Zi

Powered by blists - more mailing lists