Kernel: BPF unwind pads, THP swap, MGLRU

Kernel development on 2026-10-05 centered on memory management refinements for transparent huge pages and multi-gen LRU, alongside BPF verifier and unwind work that supports safer resource handling. Rust PCI abstractions and an x86 page-table fix rounded out the day's proposals, with one secretmem regression still under discussion.

BPF unwind exception cleanup landing pads

A v8 series of 22 patches on bpf-next implements exception cleanup landing pads that run when bpf_unwind() unwinds. The work enables Rust panic paths and orderly resource release inside BPF programs. Readers tracking BPF or Rust-in-kernel integration should watch this because it closes a gap in exception safety for eBPF.

PMD-level swap for anonymous THPs

A v8 30-patch series adds PMD-level swap entries so swapped anonymous transparent huge pages keep their huge mappings across swap-out and swap-in instead of splitting to PTEs. The change targets the memory management subsystem. It matters for workloads that rely on THP performance even when memory pressure forces swapping.

Frequency-guided MGLRU promotion

RFC v3 posts a 17-patch set that introduces frequency-guided promotion for multi-gen LRU along with flag and ABI cleanups. The authors report 10-40 percent gains across tests. Developers concerned with reclaim behavior and page promotion heuristics will find the claimed improvements relevant.

Rust PCI SR-IOV abstractions

A v4 nine-patch series adds Rust PCI SR-IOV support, including abstractions, callbacks, and VfRegistration for PF and VF drivers. The work lives in the kernel's Rust bindings for PCI. It expands the set of device features expressible in safe Rust drivers.

BPF scalar evolution for bounded loops

A large v2 43-patch series on bpf-next adds scalar evolution analysis to the BPF verifier so it can widen bounded loops and relax offset tracking. The change aims to accept more valid programs without weakening safety. Verifier improvements of this scale affect anyone writing complex BPF loops.

secretmem SIGBUS under concurrent perf

A locked-vm accounting change for secretmem produces SIGBUS on memfd_secret() pages when perf record runs concurrently. Stable trees dropped the accounting patch pending a perf fix. Users of secret memory and continuous profiling need to know the regression remains open.

Always-on mTHP support

A v8 14-patch set decouples has_transparent_hugepage() from PMD support so multi-size THP remains usable when PMD-sized pages are unavailable. The series keeps mTHP viable on configurations that lack full PMD THP. This preserves flexibility for systems that still want smaller huge-page benefits.

Incomplete x86 page-table invalidation with TCE

A single-patch fix notes that single-address invalidation is insufficient with AMD TCE and can leave stale PMD entries on pud_free_pmd_page. The change belongs to the x86 memory-management path. Correct TLB and page-table teardown on AMD platforms depends on addressing this case.