LPC 2026 sched_ext microconference advances BPF scheduler priority inversion and paravirtualized scheduling

The sched_ext microconference at Linux Plumbers Conference 2026 in Prague focused development attention on three technical fronts for the BPF-based extensible scheduler framework. Priority inversion emerged as the most discussed topic, with participants working through lock holder and lock waiter preemption mechanisms, time slice extension strategies, and the integration of proxy execution with BPF scheduler programs. Paravirtualized scheduling work aims to let host and guest kernels cooperate across VM boundaries, using sched_ext as the policy engine on the host side to coordinate scheduling decisions. Device and topology-aware scheduling discussions centered on leveraging Intel Hardware Feedback Interface (HFI), device and accelerator locality information, and cache topology data to improve task placement decisions.

These developments target kernel versions beyond the current 6.14 sched_ext merge, with no specific upstream timeline committed. The work affects x86_64 and ARM64 virtualization deployments most directly, with particular relevance for cloud providers running multi-tenant workloads and organizations deploying AI/ML infrastructure with heterogeneous accelerators. BPF scheduler authors and distribution kernel maintainers tracking sched_ext maturity will need to follow these interface evolutions closely.

Priority inversion, paravirtualization, and topology-aware scheduling in sched_ext

• Priority inversion handling: Three approaches under active discussion: lock holder and lock waiter preemption to reduce priority inversion latency, time slice extension to prevent premature descheduling of critical sections, and proxy execution integration that allows BPF schedulers to delegate scheduling authority without breaking priority chains.
• Paravirtualized scheduling: New host-guest coordination mechanism where sched_ext on the host acts as the policy engine, exposing scheduling decisions across the VM boundary rather than treating guests as opaque black boxes; requires guest-side awareness and hypervisor cooperation.
• Device and topology-aware scheduling: Integration of Intel Hardware Feedback Interface (HFI) telemetry into placement decisions, combined with device/accelerator locality awareness and cache topology data to reduce cross-NUMA or cross-cache-cluster migration penalties for accelerator-bound workloads.
• BPF scheduler interface evolution: Proxy execution and priority inversion mechanisms require new BPF program types or helper functions, creating interface stability concerns for existing out-of-tree schedulers as the subsystem matures toward broader production use.

Why it matters

Cloud providers and virtualization vendors running Linux hosts with nested or side-by-side VMs stand to gain measurable scheduling efficiency from paravirtualized sched_ext cooperation, particularly for latency-sensitive tenant workloads. AI/ML infrastructure operators with Intel-based heterogeneous compute deployments will benefit from HFI-aware placement once the topology scheduling work lands, potentially reducing memory bandwidth contention and accelerator access latency. Distribution kernel maintainers and BPF scheduler authors face ongoing interface churn as sched_ext stabilizes, requiring careful tracking of these LPC outcomes to avoid breakage in production scheduler deployments.