News › performance
By The Linux Camp News, October 4, 2026
The October 1 arm64 RFC targets CPUs where a process has run. Migration, unfinished firmware interfaces and the absence of hardware benchmarks leave the performance benefit unproven.
Arm submitted a 12-patch Linux arm64 RFC on October 1, 2026, proposing support for TLBI domains, an architecture feature intended to reduce the CPU scope of TLB invalidations on large systems. This is a request for review, not a kernel release announcement. The sources retrieved October 4 establish a development proposal, not availability in a released kernel. 2 1
TLB stands for Translation Lookaside Buffer; TLBI is the invalidation operation. FEAT_TLBID introduces domains that let an invalidation target a subset of CPUs. The proposed Linux policy is to send commands only to CPUs where a process's page tables have been installed. The intended saving therefore comes from involving fewer CPUs, rather than changing the application. This is a change in targeting scope, not a claim that every previous invalidation erased every TLB entry system-wide. 1 2
Domain descriptions come from a new TLBI ACPI table. That makes the platform description part of the proposed support, rather than this being solely a kernel-side optimization. Author Kristina Martšenko says the series uses a work-in-progress table outside the ACPI specification, with publication aimed for the end of 2026. There is no device-tree binding in this proposal because its intended platforms are primarily server-class systems. 2
Process migration can erode the intended benefit. Martšenko warns that a long-running process may eventually run on every CPU, expanding its tracked history until invalidations become broadcasts again. The operational inference is that CPU count alone would not determine the saving: how widely a workload moves across CPUs would also matter. Periodically shrinking that history is an open design question, not an implemented policy operators can tune. 2
KVM users should distinguish a possible follow-up from what this RFC implements. Martšenko proposes eventually limiting guest-related invalidations to CPUs where a guest has run, but this series only disables domains for guests. It therefore does not demonstrate the proposed guest invalidation optimization. 2
For platform evaluation, the useful next evidence is multi-domain testing and workload measurements, not a production upgrade plan. In the RFC retrieved October 4, Martšenko reports no hardware implementing the feature and no performance numbers. Testing used an Arm FVP supporting the architecture but containing only domain 0. That leaves both performance gains and behavior across multiple domains unestablished by the reported tests. 2