Linux 7.3 Networking Adds BIG TCP Overlays as AI Patch Volume Overwhelms Maintainers

Linux 7.3’s networking merge is notable for two very different reasons. Technically, it extends BIG TCP across VXLAN and Geneve UDP overlays and reduces contention around routing-rule updates. Operationally, maintainers Jakub Kicinski and Paolo Abeni say the volume of incoming work—much of it apparently generated or assisted by AI—has become overwhelming.

The biggest networking changes

BIG TCP now crosses VXLAN and Geneve overlays

The merge adds BIG TCP support for the VXLAN and Geneve drivers, allowing large TCP super-segments to remain aggregated through these UDP-based overlay tunnels. The goal is to reduce per-packet processing rather than change the configured link MTU. The reported benefit can also apply to standard-MTU deployments and systems without hardware tunnel-offload support, because software can process fewer, larger packets internally.

A supporting BIG TCP patch series describes the underlying work: tunnel paths had to stop assuming that packet lengths would remain below 64 KiB or fit in 16-bit fields. The series also covers BIG TCP IPv4 and IPv6 workloads for VXLAN and Geneve.

Less global RTNL contention for FIB rules

Another important change reduces reliance on the global RTNL lock for FIB-rule operations such as RTM_NEWRULE and RTM_DELRULE. That is aimed at systems manipulating routing rules across many network namespaces, where a global serialization point can limit parallelism.

The available evidence establishes the intended concurrency improvement, but it does not provide reliable, precise IPv4 or IPv6 benchmark speedups. Any exact percentage or multiplier would therefore go beyond the supplied evidence.

Drivers and protocol-layer updates

The networking pull also broadens hardware and protocol support:

• The wireless work includes the Morse Micro mm81x driver for long-range S1G devices and a new NXPWIFI driver.
• Preliminary Intel iXD infrastructure targets the Intel Control Plane PIC Function on Intel E2100-and-later IPUs and FNICs.
• MPTCP improves behavior under extreme memory pressure, including pruning its out-of-order receive queue.
• The wider pull includes additional wired and wireless improvements, fixes, and driver work across the networking subsystem.

These changes are part of the networking code entering the Linux 7.3 merge-window work; they should not be confused with a completed stable Linux 7.3 release.

Why the maintainers said they were “completely overwhelmed”

Kicinski and Paolo Abeni reported merging 632 net patches and 648 net-next patches—1,280 patches combined—in the networking pull. They also cautioned that this count did not capture the full review burden. Their “quick and dirty” estimate was that one-third to one-half of the net-next patches appeared to be AI-driven low-priority fixes, cleanups, or clarifications.

The pressure was part of a broader increase in submissions. In one nine-day period, the subsystem received 405 postings tagged

[PATCH net]

[PATCH net-next]

The concern is not simply that AI can produce code. AI-assisted contributors can also generate plausible explanations, bug reports, and review requests at very low cost. Human maintainers still have to determine whether each change is necessary, correct, safe, and worth merging.

How AI may be used in the review process

The maintainers said they had secured Meta-funded access and budget for multiple large language models to review patches. The immediate goal is to catch some hallucinated or weak changes before they consume more human review time.

They also plan to direct models toward routine workflow tasks, including:

• managing Patchwork queues;
• providing feedback on common submission-process problems;
• editing commit messages; and
• potentially applying patches that already carry review tags from trusted people.

That approach treats LLMs as workflow assistants rather than replacements for subsystem maintainers. The distinction matters most for concurrency-sensitive kernel code. The maintainers specifically highlighted rare-event paths such as PCIe errors and timeouts, where timing-dependent races can remain difficult to expose or prove absent. LLM review may help identify suspicious patterns, but it cannot reliably establish correctness for every uncommon interleaving.

What is still unknown

The supplied reporting does not give expected dates for Linux 7.3-rc1 or the final stable release. It describes networking changes merged during the 7.3 development cycle, not a published release timetable.

Likewise, while the RTNL changes are intended to improve parallel FIB-rule operations across network namespaces, the available material does not support a precise IPv4 or IPv6 performance figure. The defensible takeaway is architectural: less global-lock contention should give highly parallel namespace-management workloads more room to scale, but the size of the gain depends on the workload and system configuration.