Kernel RFC wave: GPU packet offload, PCS, and stable crypto fix

Kernel development on 2026-07-22 centered on networking offload and subsystem cleanups, with an RFC for in-kernel GPU packet processing and a late-stage PCS series, plus a stable backport closing a crypto race. Several memory-management and scheduler proposals also advanced toward broader architecture coverage.

In-kernel GPU network offload RFC

An RFC patch series on bpf.vger.kernel.org proposes knod, an in-kernel network offload device that would move XDP and IPsec packet processing from the host CPU onto AMD GPUs. The design embeds a GPU driver path inside the networking stack so that selected workloads can run without leaving kernel context. Readers tracking high-throughput packet paths should watch how the series balances offload gains against driver complexity and security boundaries.

fwnode PCS provider API reaches v9

A twelve-patch net-next series adds a fwnode-based PCS provider API and internal phylink handling for PCS devices. The mvpp2 maintainer gave positive review on the latest revision, indicating the abstraction is nearing merge readiness. The change matters for drivers that need firmware-node discovery of PCS blocks without hard-coded platform assumptions.

Stable fix for algif_skcipher IV race

A stable-kernel review patch backports a fix that forces synchronous processing in algif_skcipher. The original race let an attacker control the IV and reuse keystream material. Operators of long-term kernels should treat the backport as a security-relevant update for any workload using the AF_ALG skcipher interface.

binfmt_misc transparent interpreters

Christian Brauner posted a v2 series extending binfmt_misc with transparent dispatch via a T flag and BPF, plus PT_INTERP loader substitution modes. The work aims to let interpreters and loaders be swapped without userspace path rewriting. Container and compatibility runtimes are the primary audience for the new dispatch paths.

Proxy execution meets sched_ext

Version 8 of a sched_ext patchset removes the build-time exclusion that previously prevented SCHED_PROXY_EXEC and SCHED_CLASS_EXT from being enabled together. The series targets the 7.3 merge window. Extensible scheduler users who also rely on proxy execution gain a supported combination rather than a compile-time choice.

Generic set_memory core proposed

An RFC moves the x86 change_page_attr implementation into mm/ as GENERIC_SET_MEMORY and converts riscv and loongarch to the shared core. The goal is to cut duplicated page-attribute logic across architectures. MM and arch maintainers evaluating reduced per-arch surface area will find the series relevant.

Private NUMA nodes for device memory

A 36-patch v5 series introduces private NUMA nodes that isolate device memory and expose per-capability opt-ins for reclaim, migration, and related operations. The design keeps device-private pages out of the ordinary node space unless explicitly enabled. Heterogeneous-memory platforms are the intended consumers.

Ghost swapfile backend switching

An RFC for mm/swap proposes ghost swapfiles that support runtime backend switching through redirect entries, including a zswap-without-writeback mode. The mechanism lets swap targets change without tearing down the swapfile identity. Memory-constrained systems exploring flexible compressed swap will want to follow the discussion.