Linux 7.3 Networking Brings BIG TCP Overlays, Faster Routing Rules—and an AI Review Experiment

These are configuration-specific measurements. The result depends on the end-to-end setup, including gso_max_size and gro_max_size, NIC capabilities, offload settings, MTU, tunnel type, and workload. Operators should benchmark their own VXLAN or Geneve path before assuming the same percentage improvement.

Less RTNL serialization for routing-rule updates

The merge also reduces reliance on the broad RTNL lock for IPv4 and IPv6 FIB-rule insertion and deletion. Where possible, the work now uses mutex protection associated with the relevant fib_rules_ops, allowing more operations to proceed in parallel. RTNL is still required in some cases, including the first IPv4 rule’s fib_unmerge() path.

A synthetic test illustrates the potential benefit in heavily parallel namespace management. Creating 4,096 network namespaces and adding 1,024 rules to each in parallel fell from 22.752 to 0.918 seconds for IPv4—about 24.8× faster—and from 35.181 to 1.214 seconds for IPv6—about 29.0× faster. Those figures measure extreme lock contention, not a guaranteed improvement to ordinary container startup or every routing workload.

Why maintainers are turning to AI for patch triage

The networking subsystem is also becoming a test case for AI-assisted kernel development. Jakub Kicinski reported 632 net patches and 648 net-next patches, estimating that roughly one-third to one-half of net-next consisted of AI-driven, low-priority fixes, cleanups, or clarifications. That estimate corresponds to approximately 216 to 324 patches. Kicinski and Paolo Abeni described the resulting workload as “completely overwhelmed.”

The proposed response is not to replace maintainers with one model. Meta provided budget and access for multiple frontier models so patches can receive several first-pass reviews, with the goal of reducing the chance that a single model’s hallucination or mistaken interpretation reaches the process unchecked.

Planned automation includes Patchwork management, routine process feedback, commit-message editing, and potentially applying patches that trusted people have already reviewed. Those tasks are more structured than deciding whether a change is correct in a rare failure path.

That distinction is important in kernel networking. Maintainers specifically noted that unusual race conditions and hardware-recovery paths—such as PCIe errors and timeout handling—still require human scrutiny. AI can help filter repetitive work, but it does not remove the need to understand concurrency, API contracts, or recovery sequencing.

New drivers and protocol improvements

The networking pull includes several pieces of hardware and protocol enablement:

• The initial Morse Micro mm81x Wi-Fi HaLow driver supports USB and SDIO devices.
• NXPWIFI arrives as a new NXP wireless driver, while work on multi-link operation continues.
• Intel gains the initial skeleton of the iXD driver for Control Plane PIC Function support on E2100-and-later IPUs and FNICs.
• Wired networking also gains initial CXL support for AMD/Solarflare NICs and firmware-flashing support for AMD Pensando hardware.

There are smaller but useful changes beyond drivers. MPTCP gains a last-resort mechanism for pruning out-of-order queues under extreme memory pressure. This is a recovery measure for constrained conditions, not a promise of higher normal throughput.

AF_UNIX file-descriptor passing also becomes easier to diagnose. With SO_RIGHTS_NOTRUNC, a receiver can identify a rejected descriptor and its errno when an LSM blocks a particular file descriptor in an SCM_RIGHTS array, instead of losing the rest of the array at the first rejection.

When will Linux 7.3 reach users?

The first Linux 7.3 release candidate was expected around August 30, 2026, while the stable release was projected for late October, assuming the normal development schedule holds.

That does not mean most servers or desktops will receive these networking features in October. Distribution maintainers choose kernel versions, backport selected changes, run their own tests, and publish updates on separate schedules. Rolling distributions may package the kernel sooner, while fixed-release distributions can take substantially longer.

CachyOS is a plausible early adopter because it follows a rolling model, but the available August release information shows its images using Linux 7.1 rather than confirming a Linux 7.3 adoption date.

What the merge means in practice

For infrastructure teams, the most immediately relevant changes are the BIG TCP overlay path and reduced contention during large-scale network-namespace rule management. They are especially worth watching in environments that combine VXLAN or Geneve with high-throughput traffic, or that create and configure many namespaces concurrently.

For kernel developers, the AI-review experiment may be just as consequential. The merge suggests a practical division of labor: models can absorb repetitive triage and procedural feedback, while maintainers retain responsibility for the uncommon cases where a superficially plausible patch can deadlock a system, mishandle recovery, or violate an implicit synchronization rule.

The result is a substantial networking feature set heading into Linux 7.3—but one that still needs release-candidate testing, hardware validation, and distribution integration before most users can rely on it.