Re: [PATCH 1/9] mm: introduce hw_pte_t for PTE table storage

On Fri, Aug 07, 2026 at 04:24:00PM +0100, Muhammad Usama Anjum wrote:

On 07/08/2026 8:09 am, Alexander Gordeev wrote:

quoted

On Thu, Aug 06, 2026 at 09:38:39AM +0100, Muhammad Usama Anjum wrote:

quoted

pte_t is used both for logical PTE values and for entries stored in a PTE
table, so pte_t * does not distinguish a pointer to a copied value from a
pointer to table storage.

Introduce hw_pte_t as the generic name for a PTE table element. Define it
as a macro alias of pte_t by default. When an architecture selects
ARCH_HAS_HW_PTE_T, define it as a structure containing a pte_t instead.
This preserves the representation while allowing converted architectures
to enforce the distinction at compile time.

Keep the C type definitions behind an __ASSEMBLY__ check because
architecture assembly sources can include this header indirectly. Include
asm/page.h so consumers such as linux/vmalloc.h retain the page definitions
they previously obtained from that header.

Signed-off-by: Muhammad Usama Anjum <redacted>
---
Changes since RFC v1:
- Add the ARCH_HAS_HW_PTE_T opt-in and generic wrapper definition.
- Exclude the C type definitions from assembly sources.
- Update the description for the new opt-in model.
---
 MAINTAINERS                   |  1 +
 include/linux/pgtable_types.h | 17 +++++++++++++++++
 mm/Kconfig                    |  3 +++
 3 files changed, 21 insertions(+)
 create mode 100644 include/linux/pgtable_types.h

diff --git a/MAINTAINERS b/MAINTAINERS
index e9c8567308a75..7169bea968cf5 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS

@@ -16982,6 +16982,7 @@ F:	include/linux/mmu_notifier.h
 F:	include/linux/pagewalk.h
 F:	include/linux/pgalloc.h
 F:	include/linux/pgtable.h
+F:	include/linux/pgtable_types.h
 F:	include/linux/ptdump.h
 F:	include/linux/vmpressure.h
 F:	include/linux/vmstat.h

diff --git a/include/linux/pgtable_types.h b/include/linux/pgtable_types.h
new file mode 100644
index 0000000000000..70c3edd00a01b
--- /dev/null
+++ b/include/linux/pgtable_types.h

@@ -0,0 +1,17 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _LINUX_PGTABLE_TYPES_H
+#define _LINUX_PGTABLE_TYPES_H
+
+#include <asm/page.h>
+
+#ifndef __ASSEMBLY__
+
+#ifdef CONFIG_ARCH_HAS_HW_PTE_T
+typedef struct { pte_t __pte; } hw_pte_t;

On s390 it fails to compile once we do typedef hw_pte_t *pgtable_t
in asm/page.h. m68k, powerpc and sparc may also have such problem.

The below declaration helps to resolve it using forward declaration
and without meddling with headers, though I do not like it much:

typedef struct __hw_pte_t { pte_t __pte; } hw_pte_t;

Thank you for testing it out on s390.

As __hw_pte_t isn't being used yet in this series, would s390 enablement
patches add __hw_pte_t to this definition?

I hope there is a better solution. As I noted m68k, powerpc and sparc
may also be affected, so I would suggest to look into those as well.
I would prefer s390 to use the generic one rather than circumvent a
compile error in a custom way.

This could have been avoided if each architecture defined its own hw_pte_t.
But for now we are keeping the generic definition of hw_pte_t.

quoted

quoted

+#else
+#define hw_pte_t pte_t
+#endif
+
+#endif /* !__ASSEMBLY__ */
+
+#endif /* _LINUX_PGTABLE_TYPES_H */

diff --git a/mm/Kconfig b/mm/Kconfig
index 331daf7fcfab5..31ba9ebf4aafd 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig

@@ -1316,6 +1316,9 @@ comment "GUP_TEST needs to have DEBUG_FS enabled"
 config GUP_GET_PXX_LOW_HIGH
 	bool
 
+config ARCH_HAS_HW_PTE_T
+	bool
+
 config DMAPOOL_TEST
 	tristate "Enable a module to run time tests on dma_pool"
 	depends on HAS_DMA

-- 
2.47.3

-- 
Thanks,
Usama