A six-patch series from Arnd Bergmann proposes new defaults for the memory split of 32-bit kernels on Arm, x86, PowerPC and MicroBlaze, so that most boards with 1 GiB or 2 GiB of RAM would no longer need highmem. On Arm, the series would also remove highmem for ARMv4 and ARMv5 processors, remove CONFIG_HIGHPTE and add a CONFIG_VMSPLIT_2G_OPT option. Version 2 was posted on 12 November 2026 and is under review. It is not in mainline or in linux-next, and no target release has been stated.
Status on 16 November 2026: v2, under review, acked by memory-management reviewers, not applied to any maintainer tree. Not merged.
The memory split is the Kconfig choice that divides the 4 GiB virtual address space of a 32-bit kernel between user space and the kernel. This article explains what the series would change in that choice, what reviewers found, and what a team with a 32-bit Arm product should check.
The memory split on 32-bit Arm today
The kernel maps RAM permanently only inside its own part of the address space. RAM mapped in this way is called lowmem. RAM that does not fit is highmem, and the kernel must map each highmem page temporarily before it can read or write the page.
In arch/arm/Kconfig at v7.4-rc3, the choice with the prompt “Memory split” has four entries, and the default is VMSPLIT_3G:
VMSPLIT_3G PAGE_OFFSET 0xC0000000 (default)
VMSPLIT_3G_OPT PAGE_OFFSET 0xB0000000 (not available with ARM_LPAE)
VMSPLIT_2G PAGE_OFFSET 0x80000000
VMSPLIT_1G PAGE_OFFSET 0x40000000With the default, user space has 3 GiB and the kernel has 1 GiB. The commit message of patch 5 states that this upper 1 GiB is shared between lowmem, at most 896 MiB, and the vmalloc area. A board with 2 GiB of RAM therefore has more than half of its memory outside lowmem.
Product teams handle this in one of two ways. The first is CONFIG_HIGHMEM=y. For example, imx_v6_v7_defconfig sets it, and ARCH_OMAP2PLUS_TYPICAL selects it, which is how multi_v7_defconfig receives it. The second is to choose VMSPLIT_2G manually. CONFIG_HIGHPTE, which defaults to y with highmem, places user page tables in highmem.
Three separate changes affect 32-bit Arm, and they should not be confused. The first is the deprecation of legacy platforms in Linux 7.3: board files that predate device tree, Cortex-M and ARM1136r0 cores, and OABI user space. Our earlier digest on 32-bit Arm deprecations in Linux 7.3 covers it. The second is highmem. On 11 November 2026, in an unrelated discussion about splice, Linus Torvalds wrote that he wants to remove highmem entirely in the not too distant future, and that users who still need it could stay on the current longterm kernel. Arnd Bergmann posted this series the next day.
The third is what remains. No message that we read proposes the removal of 32-bit Arm itself. The commit message of patch 5 states that the i.MX6Q and Keystone2 families have expected support cycles past 2035. This series concerns such ARMv7 products.
What the series proposes for the memory split
The cover letter describes three stages. Stage 1 moves users to a larger lowmem zone through the CONFIG_VMSPLIT_* options. Stages 2 and 3 would later remove most users of highmem inside the kernel. This series implements stage 1 only.
For Arm, the six patches would do the following:
- Patch 1 adds
depends on !CPU_CACHE_VIVTtoCONFIG_HIGHMEM. ARMv4 and ARMv5 processors use VIVT caches, so these processors would lose highmem. - Patch 2 removes
ARCH_NEEDS_KMAP_HIGH_GETfrommm/highmem.c, because Arm was its last user. - Patches 3 and 4 remove
CONFIG_HIGHPTEfrom Arm and then from the generic page table code. - Patch 5 changes the default memory split and hides
VMSPLIT_3GandVMSPLIT_3G_OPTbehinddepends on !HIGHMEM || EXPERT. - Patch 6 adds
VMSPLIT_2G_OPT, withPAGE_OFFSETat 0x70000000, for a full 2 GiB of lowmem without LPAE.
After patches 5 and 6, the Arm choice would read as follows. The lines are copied from patch 6:
choice
prompt "Memory split"
depends on MMU
- default VMSPLIT_2G if HIGHMEM || ARM_LPAE
+ default VMSPLIT_2G if ARM_LPAE
+ default VMSPLIT_2G_OPT if HIGHMEM
default VMSPLIT_3G_OPT
...
+ config VMSPLIT_2G_OPT
+ depends on !ARM_LPAE
+ bool "2G/2G user/kernel split (for full 2G low memory)"The result would be three defaults. A build without highmem would use VMSPLIT_3G_OPT. A build with highmem and without LPAE would use VMSPLIT_2G_OPT. A build with LPAE would use VMSPLIT_2G, because LPAE requires a split on a gigabyte boundary.
The cost is user address space. The PAGE_OFFSET values give the size: user space would have 2816 MiB instead of 3072 MiB with VMSPLIT_3G_OPT, and 1792 MiB with VMSPLIT_2G_OPT. The commit message of patch 5 says directly: “There are likely going to be regressions from the changed defaults”. It names two sources: device drivers that fail to set the correct DMA mask, and applications that need a large virtual address space.
The author tested with QEMU guests and states that he has no hardware that still needs highmem. He estimates, without a measurement, that the new defaults would place all RAM in lowmem for more than 98% of maintained 32-bit embedded systems.
The series can be fetched with b4, or from the branch that the cover letter names:
raghu@techveda.org:~$ b4 am 20261112112528.1116405-1-arnd@kernel.org
raghu@techveda.org:~$ git fetch https://git.kernel.org/pub/scm/linux/kernel/git/arnd/playground.git highmem-7.4The review
Version 1 had four patches and was posted on 19 December 2025. Dave Hansen of Intel acked the x86 part and asked for simpler defaults. His argument was that a user who upgrades keeps the existing configuration and is not affected by a new default. Arnd Bergmann disagreed about the effect. He replied that he expects the split to change for projects that store a defconfig and for users with a full .config, unless CONFIG_EXPERT is set. Christophe Leroy tested on PowerPC and reported two problems: the e500 port needed a larger CONFIG_LOWMEM_CAM_NUM, and the KASAN shadow area overlapped the enlarged lowmem.
Version 2 adds the two CONFIG_HIGHPTE patches, which David Hildenbrand had asked for on 12 November. Mike Rapoport acked patches 2 to 5 and listed remaining references to CONFIG_HIGHPTE. David Hildenbrand acked patch 4 on the condition that these are removed. Lorenzo Stoakes acked all six patches on 13 November, three of them conditionally.
The findings with the most effect on products came from an automated review, sent by sashiko-bot on 13 November. It listed nine possible problems in patch 5. The author accepted the Arm findings:
- Patch 5 removed
select HIGHMEMfromARCH_OMAP2PLUS_TYPICAL. This would have disabled highmem inmulti_v7_defconfigandomap2plus_defconfigand limited boards to 1 GiB of RAM. The author wrote that this was not intended and withdrew the hunk. He proposed a separate patch that setsCONFIG_VMSPLIT_2G_OPT=yin both defconfigs, and expects it to be controversial. - Patch 5 removed
CONFIG_HIGHMEM=yfromaspeed_g5_defconfig, which would have truncated RAM on AST2600 boards with 2 GiB. The author withdrew this hunk. pxa3xx_cpu_pm_suspend()inarch/arm/mach-pxa/pxa3xx.cuses the address 0xc0000000 as a constant. With a differentPAGE_OFFSET, suspend would write to the wrong address. The author added a fix that usesPAGE_OFFSET.
The bot also repeated both PowerPC problems from version 1. The author answered with proposed changes and wrote that he suspects more code assumes 0xc0000000. H. Peter Anvin recalled that this had been tried once before and reverted. He gave no reference.
Where it stands
On 16 November 2026, version 2 is the latest version on the list. The series carries acks from Mike Rapoport, David Hildenbrand and Lorenzo Stoakes. No maintainer has stated that it is applied. In linux-next, arch/arm/Kconfig at next-20261113 does not contain VMSPLIT_2G_OPT, and the same is true of mainline at v7.4-rc3. No message in the thread names a target release.
The author’s branch contained ten commits when we read it, including the PXA fix and a defconfig commit for VMSPLIT_2G_OPT. David Hildenbrand wrote that he assumes a version 3 will follow. The next steps are that version 3, tests on PowerPC hardware, and a decision on multi_v7_defconfig.
What it would mean for a product team
No action is needed on a shipped kernel. The series changes nothing until it is merged. Three checks are useful now.
First, record the current configuration of each 32-bit product:
root@imx6qsabresd:~# zcat /proc/config.gz | grep -E 'VMSPLIT|HIGHMEM|HIGHPTE|ARM_LPAE|CONFIG_EXPERT'
root@imx6qsabresd:~# grep -E 'HighTotal|LowTotal' /proc/meminfoA product with CONFIG_HIGHMEM=y, the default VMSPLIT_3G and no CONFIG_EXPERT is the case that the proposal would move. As posted, the same configuration would select VMSPLIT_2G_OPT after an upgrade, and user space would have 1792 MiB of address space instead of 3 GiB. A team that needs the present layout would have to set CONFIG_EXPERT=y and select the memory split explicitly.
Second, test the product with a 2G/2G memory split before an upgrade selects one. VMSPLIT_2G exists today and gives 2 GiB of user address space, close to the proposed option. Build the current kernel with it, then run the applications that map the most memory. The author reports that Firefox 140 on Debian armhf still displayed complex web sites with VMSPLIT_2G_OPT. That is one application in a QEMU guest.
Third, search out-of-tree drivers and board code for the constant 0xc0000000 and for missing DMA mask calls.
The series would not help a board with more than 2 GiB of RAM, which would still need highmem. It also would not remove highmem; that is the subject of the later stages.
What was checked and what could not be verified
We read the cover letter, the six patches and all 29 messages of the version 2 thread, the version 1 cover letter and its replies, and the messages of 11 and 12 November that preceded the new posting. We read the Arm Kconfig files, pxa3xx.c and four defconfig files at v7.4-rc3, and arch/arm/Kconfig at next-20261113.
We did not build or boot any kernel with the series. The address space figures are calculated from the PAGE_OFFSET values in the patches. The statement about a saved configuration follows from the Kconfig dependencies and from the author’s reply in version 1; we did not run make olddefconfig to confirm it. We read only the titles of the commits on the author’s branch, so the text of version 3 is unknown.
Key takeaways
- The series is a proposal. It is not in mainline at v7.4-rc3 or in next-20261113, and no release has been named.
- As posted, a 32-bit Arm build with highmem and without
CONFIG_EXPERTwould change fromVMSPLIT_3GtoVMSPLIT_2G_OPT. - The proposed memory split gives lowmem for 2 GiB of RAM and leaves 1792 MiB of address space for each process.
- The series does not remove 32-bit Arm. It prepares ARMv7 products for a kernel with less highmem.
- Review found constants that assume a
PAGE_OFFSETof 0xc0000000. Out-of-tree code can contain the same assumption.
Frequently asked questions
Has the default memory split on 32-bit Arm changed?
No. At v7.4-rc3 the default is VMSPLIT_3G. The change is proposed in version 2 of a series posted on 12 November 2026, which is under review and not merged.
Would a product be able to keep the 3G/1G memory split with highmem?
As posted, yes. VMSPLIT_3G and VMSPLIT_3G_OPT would remain selectable with highmem when CONFIG_EXPERT=y is set.
In which kernel release would the change arrive?
No target release has been stated. The branch name highmem-7.4 shows the base of the patches, which is 7.4-rc2.
References
- [PATCH v2 0/6] mm: highmem cleanups — Arnd Bergmann, linux-mm, 12 November 2026 (cover letter and thread)
- [PATCH 0/4] mm: increase lowmem size in linux-7.0 — version 1, 19 December 2025
- Linus Torvalds on highmem, 11 November 2026 (splice thread)
- Author’s branch highmem-7.4, arnd/playground.git
- arch/arm/Kconfig at v7.4-rc3
- arch/arm/Kconfig in linux-next (next-20261113)
- LWN: summary of the highmem session at the 2025 kernel summit (linked from the cover letter)




