Linux Kernel Prepares Massive Purge of Legacy 32-Bit ARM Code

Arnd Bergmann has spent years steering the ARM side of the Linux kernel. Now his latest work promises to strip away a quarter million lines of code that almost no one uses anymore. The change marks another step in a long pattern of shedding support for hardware that has faded from relevance.

Just days ago Phoronix reported on Bergmann’s board-remove-7.3 Git branch. The collection of roughly 300 patches deletes some 247,000 lines while adding back only about 5,000. That net reduction equals roughly 0.6 percent of the entire kernel source tree, which now sits near 41 million lines.

The removals follow directly from deprecations that landed in Linux 7.3. Those patches already cut around 55,000 lines tied to the platform code itself. Yet the real savings come from drivers that existed solely to support those long-outdated systems. Once the platforms go, the drivers become orphans. And orphans get deleted.

Platforms on the list include SA1100, Footbridge, RISCPC, Orion, Dove, MV78xx0, OMAP24xx, i.MX31, i.MX targets without an MMU, LPC18xx, STM32f4/f7/h7 MCUs, Versatile MPS2, AT91 SAMV7, Axxia, and legacy PXA board files. Many of these names evoke images of early 2000s embedded boards, industrial controllers, and hobbyist projects that have not seen a mainline kernel in years.

Bergmann first floated a broad deprecation plan back in 2024. He admitted in a later mailing list post that he “dropped the ball” on following through. The renewed push aligns with the end of this year’s long-term support kernel. He now eyes Linux 7.5 for the final cut to keep dependency chains clean. Some patches could land in 7.4. Most observers expect the later target.

This effort does not stand alone. The kernel has seen similar cleanups before. In early 2023 Bergmann drove the removal of Samsung s3c24xx code and various old board support files, trimming 154,000 lines. Earlier this year Linus Torvalds merged a networking pull request that erased the entire ISDN subsystem, several ancient Ethernet drivers, and amateur radio support. That change removed 138,000 lines and was motivated in part by the burden of AI-generated bug reports against code no one maintains.

So the ARM purge fits a larger trend. Maintainers grow tired of carrying dead weight. Modern development focuses on device trees, upstream first silicon from Apple, Qualcomm, and others, and 64-bit architectures that dominate new designs. The 32-bit ARM platforms targeted here predate widespread device tree adoption. Their board files relied on static machine descriptions that have become unnecessary complexity.

Yet the decision carries trade-offs. A handful of surviving board files for TI OMAP1 and Samsung S3C64xx parts remain because they still have known users. Cortex-M microcontroller support sees partial cuts. Features such as iWMMXt floating point and the original ARM1136r0 core also face elimination. Big-endian ARMv7 stays marked broken for now but avoids outright deletion.

Kernel size matters less than it once did. Storage and memory have grown cheap. The real value lies in reduced maintenance burden, faster compile times, and fewer places where obscure bugs can hide. Bergmann’s patches also remove documentation, device tree bindings, and defconfigs that no longer serve any purpose.

Critics might argue that someone, somewhere still runs a StrongARM-based system or an old RISC PC with a mainline kernel. The probability remains low. As one recent analysis noted, the chance of anyone booting a current upstream release on these platforms in 2026 approaches negligible. Those users have long since moved to out-of-tree forks, older kernels, or entirely different operating systems.

The broader context includes parallel x86 cleanups. Support for the Intel 486 and certain early Pentium-era chips has already been dropped. Russia’s Baikal CPUs saw removal. Each decision reflects the same calculation: finite maintainer time should go toward hardware that ships today.

Bergmann’s work stands out for its scale. Three hundred patches represent a substantial series. Each one must be reviewed, tested where possible, and integrated without breaking the remaining ARM32 support. The git branch already exists for inspection. Interested parties can examine the exact deletions before they reach mainline.

Discussions on the Linux Kernel Mailing List have drawn little objection so far. That silence speaks volumes. When similar proposals surfaced years ago, a few voices defended niche platforms. This time the consensus appears to accept the inevitable. The code has simply aged out.

Future kernels will compile faster on ARM64 hosts. Binary sizes will shrink slightly. Security auditors will have fewer legacy code paths to worry about. And the ARM subsystem can focus more sharply on new silicon from Arm Ltd., its partners, and the growing embedded AI segment.

None of this erases history. The deleted code powered early smartphones, set-top boxes, network appliances, and countless custom boards. It served its time. Now the kernel moves on. Bergmann and his colleagues have performed this kind of housekeeping before. They will likely do it again as other corners of the tree accumulate their own obsolete layers.

The Linux project has always balanced innovation against compatibility. In this case the balance tips clearly toward simplification. A quarter million lines represent real effort saved. That effort can now shift to features that matter for the next decade of devices rather than the last.