KVM CCA, memory safety, and mm subsystem churn

Kernel traffic on 2026-08-13 was dominated by large memory and virtualization series, alongside security fixes and a late GPU scheduler regression. Confidential computing, NUMA isolation, and BPF diagnostics drew the bulk of review attention.

Arm CCA support for KVM reaches v16

A 45-patch series (v16) continues work to add Arm Confidential Compute Architecture support to KVM, covering confidential VMs and the Realm Management Monitor. Reviewers and testers exchanged hardware feedback across 86 messages. The work matters for anyone tracking hardware-backed confidential VMs on arm64.

hp-bioscfg memory safety fixes

A nine-patch v2 series for the platform/x86 hp-bioscfg driver addresses heap out-of-bounds reads and writes, off-by-one errors, and ACPI parsing bugs. The changes are marked for stable. Operators of HP systems that use the driver should treat the series as a concrete safety cleanup.

drm/sched FAIR policy regression on 9070XT

A regression report states that the FAIR scheduling policy in drm/sched causes serious performance loss at max GPU load on 9070XT hardware. Maintainers discussed a revert days before the 7.2 release. GPU users and drm contributors have a short window to weigh the tradeoff before freeze.

BPF verifier error redesign

A 16-patch v4 series on bpf-next redesigns verification error reporting so the verifier can emit source annotations, causal history, and categorized safety reports. Three participants iterated on the design. Clearer diagnostics reduce the cost of debugging rejected programs for BPF authors and tooling.

pidfd namespace ioctl locking for CVE-2026-64371

A patch holds exec_update_lock around pidfd namespace ioctls so behavior matches the earlier procfs fix for CVE-2026-64371. Four participants reviewed the change. The fix closes a locking gap that could otherwise leave namespace operations exposed during exec.

Private Memory NUMA nodes

A 36-patch v5 series introduces N_MEMORY_PRIVATE nodes so device memory can be default-isolated while still offering opt-in mm services. Nine participants discussed the design over 64 messages. The approach targets accelerator and device-memory topologies that should not join ordinary page allocation by default.

Hot page tracking versus existing tiering

Maintainers debated an eight-patch v8 hot-page promotion subsystem (pghot) aimed at CXL tiering, comparing it with DAMON and existing NUMA balancing. The thread ran hot across 40 messages. The outcome will shape how tiered memory promotion is centralized or left to existing mechanisms.

ALLOC_UNMAPPED and AS_NO_DIRECT_MAP

A 26-patch v3 series adds ALLOC_UNMAPPED and direct-map management hooks for guest_memfd and ASI use cases. Six participants reviewed the mm changes. The work supports memory that must stay off the direct map for confidential or isolated guest scenarios.