Kernel: in-tree QUIC, XFS recovery bugs, BPF and build proposals

Kernel lists carried several large infrastructure series, a multi-bug XFS log-recovery report, and a long argument over AI attribution tags. Networking, the BPF verifier, kbuild, the page cache, and device binding all saw concrete patch traffic.

In-kernel QUIC infrastructure reaches v18

A 15-patch net-next series proposes in-kernel QUIC infrastructure and core subcomponents for kernel consumers. It includes a socket API plus connection-ID, stream, crypto, timer, and packet handling. The series is still a proposal, but it would put a QUIC implementation inside the kernel rather than leaving the protocol entirely to userspace stacks.

XFS log recovery can Oops on mount

A reporter describes a null-pointer Oops during XFS mount log recovery: XLOG_WAS_CONT_TRANS on a transaction with no buffered regions reads ri_buf[-1]. The same report details nine distinct out-of-bounds, null-deref, and unvalidated-geometry bugs in XFS log recovery. Mount-time recovery of a damaged or untrusted log is on the path that turns a filesystem image into a running system, so these are stability and safety bugs, not cosmetic ones.

BPF verifier drops scalar IDs for scalar evolution

Alexei posted a 26-patch bpf-next series that finds linked registers by scalar evolution and removes scalar IDs. The change is aimed at linked-register tracking and counted loops in the verifier. Verifier precision is what decides which BPF programs are accepted, so a rewrite of that tracking model is a core tooling change, not a peripheral optimization.

kbuild series claims large build speedups

A v5 series of 11 patches proposes kbuild, modpost, and objtool changes that claim 36 to 90 percent faster allmodconfig, incremental, and no-op builds. If those numbers survive review, they would cut the cost of the builds kernel developers run most often.

BPF kthreads and waitqueues

A 20-patch bpf-next series adds bpf_task_park and wake helpers together with BPF kthreads. The stated goal is sleepable programs and kernel threads controlled by BPF. That would extend BPF from short hooks toward longer-lived kernel work that can sleep.

Lockless page-cache reads without folio references

An RFC v2, originated by Linus and posted from Meta, adds PG_rcu_free plus an RCU grace period so filemap and shmem reads can proceed without folio references. The aim is to avoid folio refcount contention. It is still an RFC, so the design of lockless cache reads is open, not settled.

Device trust checked before driver binding

A 17-patch series generalizes the PCI untrusted flag into cross-bus device trust levels (disabled, adversary, and full) with userspace policy hooks before driver binding. The stated setting is confidential computing. Binding policy would then apply across buses, not only to PCI devices marked untrusted.

Patch to drop Assisted-by tags draws a large thread

A documentation patch to remove Assisted-by references drew 83 messages from 21 participants. The discussion covers AI patch volume, maintainer burnout, and review process. The change itself is a tag in the docs, but the thread shows the community has not agreed how AI-assisted submissions should be marked or reviewed.