Linux arm64 set to drop big-endian kernel support and boot CPUs in parallel

Arm64 sheds big-endian kernel support

The direction of Linux on 64-bit Arm was set this week by two patch series posted by arm64 maintainer Will Deacon on the same day. According to dev.to, the first is a second version of a 14-patch set, posted on 7 September, that removes around 540 lines of big-endian handling from the arm64 tree. The BPF JIT, crypto assembly, optimised string routines, the compat vDSO build and the KVM hypervisor object all lose their big-endian paths, and CONFIG_CPU_BIG_ENDIAN disappears from arch/arm64 entirely. The v2 series, rebased onto 7.3-rc1, now forces little-endian explicitly during early boot, and it already carries acknowledgements from maintainers Catalin Marinas, Arnd Bergmann, Eric Biggers and Marc Zyngier. The deletion is planned for the 7.4 merge window.

As dev.to explains, this is the end of a staged process rather than a snap decision: the big-endian configuration was made to depend on BROKEN in Linux 6.18, which made it unbuildable in practice, and dead code has been leaving the tree since. Big-endian arm64 hardware exists mostly in networking and telecom appliances running vendor kernels, and mainline never really had a big-endian arm64 user base. One detail matters for embedded teams: SETEND emulation for 32-bit user tasks stays, so mixed-endian 32-bit userspace keeps working. Bergmann also confirmed in review that 32-bit Arm is on the same track, with big-endian marked deprecated in 7.3 and the ixp4xx platform repaired to run little-endian; the arm32 removal is expected to wait one more long-term release.

Parallel CPU onlining through PSCI

The second series, 19 patches, brings parallel secondary CPU bringup to arm64 by wiring it into the generic HOTPLUG_PARALLEL machinery that x86, MIPS and RISC-V already use — arm64 has lacked it until now. It supersedes earlier HOTPLUG_PARALLEL work posted by Jinjie Ruan of Huawei in June, whose cooperation the cover letter credits.

The mechanism centres on PSCI v0.2, which gave the CPU_ON call a context argument that firmware hands to the incoming CPU. The series uses it to pass each secondary CPU its boot parameters directly through firmware, removing the serialisation point where every CPU had to identify itself against shared global state. With that gone, arm64 can drop its custom secondary-CPU synchronisation and adopt the shared parallel, split-startup and core-sync code, while a final patch hardens bringup against broken PSCI firmware on shipping boards.

The performance figures come from the predecessor series: Ruan measured a 192-core HiSilicon server bringing its CPUs up in 8.6 seconds instead of 14.6, a 41 percent reduction, with cuts of 62 to 65 percent on 32-core and 64-core boards. There is also a security benefit inside confidential guests, where the new scheme defends against a malicious hypervisor injecting secondary CPUs after the onlining window has closed. Deacon presented the design at last year's KVM Forum, and since the series is based on 7.3-rc2, the earliest realistic merge target is 7.4.

A busy 7.3-rc2, Buildroot 2026.08 and crawler load

Linus Torvalds released Linux 7.3-rc2 on 6 September and described it as an unusually busy release at what is normally the calmest point of the cycle, quipping that people will blame AI for the volume either way. The fixes span filesystems, networking, BPF, a large DRM pull and late EDAC updates, with scheduler corrections for cache-aware load balancing on hybrid CPUs leading the way.

In the embedded world, Buildroot 2026.08 shipped with close to 1,000 changes, including Linux 7.1.x headers, GCC 16.2.0, glibc 2.44, M68K nommu support and IBM Power 10/11 support. Kernel.org also published first-party measurements showing AI crawlers consume roughly a fifth of its git infrastructure capacity — about 6 million daily requests, with only around 2 percent of traffic deemed legitimate.

Why it matters

Both arm64 series change what the platform looks like for the next decade. One deletes a configuration axis that almost nobody on mainline exercised, leaving less conditional assembly in hot paths; the other modernises how PSCI-based machines boot, with gains that scale with core count and land squarely in the fast-boot windows products care about. Teams that genuinely need a big-endian arm64 kernel now have a hard ceiling at 7.3 and should open migration talks with their silicon vendor. Everyone else benefits from simpler code and faster bringup, but the parallel path will stress PSCI implementations in vendor firmware that were only ever exercised serially — maintainers of Trusted Firmware-A ports are the natural early testers here.