Kernel lock fixes, VFS open flags, BPF safety

Kernel traffic centered on lock-hold reductions in driver core and kernfs, plus VFS and BPF hardening. Maintainers also debated LLM-assisted commit norms and new agent tooling for Fixes and stable tags.

Driver core drops device lock in supplier sysfs show

A v1 patch stops waiting_for_supplier_show() from taking the device lock. The change fixes a deadlock when that sysfs attribute is read during a successful probe, a hazard present since 5.9. Device-model and probe-path reliability depend on the correction.

VFS O_CREAT|O_DIRECTORY for atomic mkdir and open

Version 7 of a 12-part series adds O_CREAT|O_DIRECTORY semantics to open*(2) so mkdir and open can complete atomically with a pinned file descriptor. The set also touches namei and kernfs. Callers that need race-free directory creation will care about the new flag combination.

Kernfs allocates inodes outside kernfs_rwsem

A v3 patch moves kernfs inode and dentry allocation off the kernfs_rwsem path. Meta reported long cgroupfs stalls when the lock was held across memcg-reclaim allocations. The work targets latency under heavy cgroup activity.

Agent skills for Fixes, stable tags, and commit text

Sasha Levin posted v3 of a seven-patch series that teaches LLM agents to handle Fixes tags, stable tags, similar-bug search, and commit messages. Greg KH and others reviewed. Maintainers weighing automated patch hygiene will follow the thread.

Docs remove Assisted-by references

A documentation patch drops Assisted-by references after the Maintainers Summit. A long discussion among many maintainers weighed disclosure rules and how to detect LLM-generated contributions. The outcome shapes attribution practice for AI-assisted kernel work.

BPF memory-safety, lifetime, and locking followups

A 16-patch bpf-next v3 series fixes memory-safety, lifetime, and locking bugs across BPF programs, dynptrs, cpumap, and the verifier. The set continues earlier hardening. BPF and verifier developers should track the corrections.

mremap stops double-charging shrinking moves

A v1 mm patch prevents mremap() from charging a shrinking move twice under MAYMOVE|FIXED. The bug is a v6.15 regression that leaked vm_committed_as. Correct committed-memory accounting is the reason to apply it.

BPF subprog unwinding via r3

Alexei posted patches that unwind BPF subprograms through register r3, with x86 and arm64 JIT, verifier, and selftest work. The design is framed as a lighter alternative to table-driven unwinding. JIT and verifier maintainers comparing unwind costs will weigh the approach.