Kernel MM patches, NUMA nodes, and low-level fixes

Memory management work led the day, with series on multi-size THP enablement, private NUMA nodes, reclaim storms, and NUMA balancing for shmem. Security and stability items included a ptdump UAF fix, a btrfs folio lock leak, virtio validation, and merged IPI preemption changes.

Always-enable multi-size THP support

A v6 patch series for the mm subsystem introduces pgtable_has_pmd_leaves() to decouple multi-size transparent huge pages (mTHP) from PMD-sized THP. The change aims to enable mTHP support even when PMD support is absent. Developers tracking huge-page paths should watch how this broadens mTHP availability.

NUMA balancing restored for shmem

A patch fixes a long-standing bug that disabled automatic NUMA balancing on shmem. mm maintainers discussed KVM side effects and related documentation. The fix matters for workloads that rely on balanced shared-memory placement across nodes.

Reclaim storms under defrag_mode

A v2 patchset addresses reclaim and swap storms caused by vm.defrag_mode=1. It makes non-movable allocations participate in pageblock-order compaction and reclaim. Systems that enable defrag_mode may see more predictable reclaim behavior once the series lands.

Private Memory NUMA Nodes

The v5 series introduces N_MEMORY_PRIVATE nodes for default-isolated device memory with opt-in mm services. The design keeps such memory out of general allocation unless services explicitly join. Heterogeneous-memory and device-driver work stands to gain a clearer isolation model.

btrfs folio lock leak in 7.2-rc1

A regression report identifies a folio lock leak in writepage_delalloc() for large data folios dirtied via GUP without subpage bits. A proposed fix is under discussion among the participants. btrfs users on 7.2-rc kernels should track the resolution to avoid lock leaks under that dirtying path.

UAF race in ptdump versus vmap freeing

An mm-hotfixes v6 series fixes a use-after-free race between ptdump and vmap or CPA kernel page-table freeing. The change closes a concurrency hole in page-table teardown. Anyone using page-table dump facilities or related debug paths has a direct interest in the fix.

Preemption during IPI completion waits

A v10 series that allows preemption while waiting for IPI completion has been merged into the tip smp/core and x86/mm branches. The work targets improved real-time performance by reducing non-preemptible wait windows. Latency-sensitive and real-time configurations are the primary beneficiaries.

Virtio device-value validation

A four-patch set adds validation of device-reported values across virtio drivers and includes a UAF fix touching virtio-mem, input, and console. Discussion covered the underlying trust model and the use of LLM-assisted checks. Virtio driver authors and anyone exposing untrusted devices should review the validation approach.