Routing-rule scalability: IPv4/IPv6 FIB-rule add/delete operations were moved away from broad RTNL serialization toward per-fib_rules_ops mutex protection where possible. RTNL remains necessary for some cases, including the first IPv4 rule’s fib_unmerge() path.

• A synthetic test creating 4,096 namespaces and inserting 1,024 rules into each in parallel fell from 22.752 to 0.918 seconds for IPv4—about 24.8×—and from 35.181 to 1.214 seconds for IPv6—about 29.0×. That measures extreme parallel namespace contention, not normal container-startup performance.

AI patch flood: Networking maintainer Jakub Kicinski counted 632 net and 648 net-next patches, and estimated that one-third to one-half of net-next were AI-driven, low-priority fixes, cleanups, or clarifications—roughly 216–324 patches. He and Paolo Abeni described themselves as “completely overwhelmed.”

AI as a countermeasure, not a replacement: Meta supplied budget/access for multiple frontier models to review each patch, intended to reduce individual-model hallucinations. The planned automation targets Patchwork management, common process feedback, commit-message editing, and possibly applying patches already reviewed by trusted people.

• Maintainers explicitly do not see this as sufficient for rare, race-prone paths—such as PCIe errors and timeouts—where API behavior and recovery sequencing still need human scrutiny.

Drivers and protocol work:

• Wireless gained the initial Morse Micro mm81x Wi‑Fi HaLow driver, for USB and SDIO devices, and NXPWIFI as a new NXP wireless driver; MLO work also continued.
• Wired work includes the initial skeleton Intel iXD driver, intended for Intel Control Plane PIC Function support on E2100-and-later IPUs/FNICs, plus initial CXL support for AMD/Solarflare NICs and AMD Pensando firmware-flashing support.
• MPTCP now has a last-resort out-of-order-queue pruning mechanism under extreme memory pressure, aimed at recovery/retransmission rather than normal throughput.
• AF_UNIX SCM_RIGHTS gains improved reporting when an LSM rejects particular file descriptors: with SO_RIGHTS_NOTRUNC, the receiver can identify the rejected descriptor and its errno instead of losing the remainder of the array at the first failure.

Schedule and availability: The first 7.3 release candidate was expected around August 30, with a stable kernel projected for late October, subject to the normal release cycle. Most users of fixed-release distributions should expect a substantially longer wait because distributions select, backport, test, and ship kernels on their own schedules. Rolling distributions can package it earlier; CachyOS is a plausible early adopter, but there is no evidence here of a committed 7.3 adoption date. Its August image was still shipping Linux 7.1.

Routing-rule scalability: IPv4/IPv6 FIB-rule add/delete operations were moved away from broad RTNL serialization toward per-fib_rules_ops mutex protection where possible. RTNL remains necessary for some cases, including the first IPv4 rule’s fib_unmerge() path.

• A synthetic test creating 4,096 namespaces and inserting 1,024 rules into each in parallel fell from 22.752 to 0.918 seconds for IPv4—about 24.8×—and from 35.181 to 1.214 seconds for IPv6—about 29.0×. That measures extreme parallel namespace contention, not normal container-startup performance.

AI patch flood: Networking maintainer Jakub Kicinski counted 632 net and 648 net-next patches, and estimated that one-third to one-half of net-next were AI-driven, low-priority fixes, cleanups, or clarifications—roughly 216–324 patches. He and Paolo Abeni described themselves as “completely overwhelmed.”

AI as a countermeasure, not a replacement: Meta supplied budget/access for multiple frontier models to review each patch, intended to reduce individual-model hallucinations. The planned automation targets Patchwork management, common process feedback, commit-message editing, and possibly applying patches already reviewed by trusted people.

• Maintainers explicitly do not see this as sufficient for rare, race-prone paths—such as PCIe errors and timeouts—where API behavior and recovery sequencing still need human scrutiny.

Drivers and protocol work:

• Wireless gained the initial Morse Micro mm81x Wi‑Fi HaLow driver, for USB and SDIO devices, and NXPWIFI as a new NXP wireless driver; MLO work also continued.
• Wired work includes the initial skeleton Intel iXD driver, intended for Intel Control Plane PIC Function support on E2100-and-later IPUs/FNICs, plus initial CXL support for AMD/Solarflare NICs and AMD Pensando firmware-flashing support.
• MPTCP now has a last-resort out-of-order-queue pruning mechanism under extreme memory pressure, aimed at recovery/retransmission rather than normal throughput.
• AF_UNIX SCM_RIGHTS gains improved reporting when an LSM rejects particular file descriptors: with SO_RIGHTS_NOTRUNC, the receiver can identify the rejected descriptor and its errno instead of losing the remainder of the array at the first failure.

Schedule and availability: The first 7.3 release candidate was expected around August 30, with a stable kernel projected for late October, subject to the normal release cycle. Most users of fixed-release distributions should expect a substantially longer wait because distributions select, backport, test, and ship kernels on their own schedules. Rolling distributions can package it earlier; CachyOS is a plausible early adopter, but there is no evidence here of a committed 7.3 adoption date. Its August image was still shipping Linux 7.1.