Linux memory management has lost one of its most active reviewers to the AI patch flood: Lorenzo Stoakes removed his reviewer entries from five MAINTAINERS areas on 10 November, stating that AI-generated series have outgrown his review bandwidth. Linux 7.4 brings two changes embedded teams should register — dm-verity can now be built to require root-hash signatures unconditionally, and 32-bit Arm loses supersection mappings, capping non-LPAE kernels at 32-bit physical addresses. The same release adds mainline Raspberry Pi 5 RTC and IOMMU drivers. Away from the kernel tree, Renesas introduced the RZ/T2N industrial MPU and Bootlin released sbom-cve-check 1.4.0 with OpenVEX export.
This edition records the week in which the AI patch flood stopped being only a statistic and started rewriting the kernel’s MAINTAINERS file. Around it, Linux 7.4 tightens verified boot, removes a 32-bit Arm legacy mechanism and improves Raspberry Pi 5 support, Renesas adds an industrial MPU aimed at robotics, and Bootlin updates its SBOM vulnerability checker.
In this edition
- Memory management loses a reviewer to the AI patch flood — Lorenzo Stoakes removed his reviewer entries from five memory-management areas, citing the volume of AI-generated series. — process signal, read before your next upstream submission
- dm-verity signatures can now be made mandatory — Linux 7.4 adds a build option that rejects unsigned dm-verity devices and cannot be overridden from the kernel command line. — adopt for verified boot
- 32-bit Arm supersection mappings removed — non-LPAE kernels are limited to 32-bit physical addresses from Linux 7.4. — check your platform
- Raspberry Pi 5 RTC and IOMMU drivers reach mainline — two more pieces of the Pi 5 work without the downstream kernel, but not in the 2026 LTS. — long-term watch
- Renesas RZ/T2N pairs Linux with four-axis motor control — two Cortex-A55 cores beside two Cortex-R52 cores, with EtherCAT and TSN on one chip. — evaluate for industrial designs
- sbom-cve-check 1.4.0 adds OpenVEX export — the SBOM vulnerability checker also improves its Linux kernel CVE version-range analysis. — adopt in your CVE workflow
Memory management loses a reviewer to the AI patch flood
On 10 November, Lorenzo Stoakes of Arm posted a four-patch series removing his reviewer entries from five sections of the kernel’s MAINTAINERS file: memory management core, memory management miscellaneous, memory reclaim, memory management for Rust, and the Rust allocator. His maintainer entries are unchanged. The reason is stated plainly in the cover letter: given the throughput resulting from AI-generated series, he wrote, “I unfortunately lack the bandwidth to generally be able to offer review” outside the areas he maintains.
A fourth patch adds a CREDITS entry recording his time as co-maintainer of transparent huge pages, a role Zi Yan of NVIDIA now carries. Stoakes described the change as administrative rather than a departure: he expects to keep reviewing where needed, and much of his attention has moved to KVM work. The series collected acknowledgements within hours from maintainers at Arm, Meta, NVIDIA, Microsoft and the Rust-for-Linux side. David Hildenbrand, who maintains the core memory-management section, agreed that the current volume has become unmanageable and proposed that Andrew Morton queue the series directly.
The context makes this more than a personnel note. The 7.4 merge window closed at 18,591 non-merge changesets, the largest on record, and reviewers — not build capacity, not CI — are the scarce resource in the kernel process. TECH VEDA has tracked the AI patch flood through the maintainers’ own numbers, most recently in the 10 November edition. What is new here is a core reviewer reducing his formal role and naming AI-generated submission volume as the direct cause. MAINTAINERS entries are how patch routing works; when they shrink, review latency is no longer an anecdote, it is recorded in the file the tooling reads.
What it means for teams contributing to memory management
Expect longer review latency for memory-management series while the AI patch flood lasts, and plan upstream work accordingly. A product schedule that depends on an in-flight MM feature deserves added margin. Smaller series with a clear problem statement stand a better chance of attracting the remaining reviewers’ time than broad refactoring. Disclose any tool assistance honestly and test on real hardware before posting: submissions that read as machine-generated are precisely what the remaining reviewers are triaging against.
dm-verity signatures can now be made mandatory
Linux 7.4 adds a dm-verity build option, CONFIG_DM_VERITY_VERIFY_ROOTHASH_SIG_FORCE, contributed by Ayoub Zaki. Until now, a kernel with CONFIG_DM_VERITY_VERIFY_ROOTHASH_SIG could validate a PKCS#7 signature on a verity root hash, but whether a signature was required was decided at boot by the dm_verity.require_signatures module parameter, which defaults to off. The new option builds the requirement in: the parameter defaults to on and can no longer be changed, and a verity device created without a valid root-hash signature is rejected.
The gap this closes is real on embedded systems. The root hash anchors the integrity of the whole read-only filesystem, and the latest part of our RISC-V verified-boot series is about exactly that anchoring problem. On many devices, however, the kernel command line is writable territory — a U-Boot environment, an extlinux.conf, a recovery shell. On such a device, a signature requirement that lives in a module parameter is advisory. With the new option built in, dm-verity signature enforcement can no longer be switched off from the kernel command line.
The option appears in 7.4 only; no stable or longterm series carries it today. The signing key must be reachable through the kernel’s trusted keyrings, and the existing companion options for the secondary and platform keyrings apply unchanged.
How to use it
Enable the force option alongside the existing signature support, and sign the root hash when the image is built (veritysetup accepts a detached root-hash signature). Confirm what a kernel enforces before trusting it:
raghu@techveda.org:~$ zcat /proc/config.gz | grep DM_VERITY_VERIFY_ROOTHASH_SIG
CONFIG_DM_VERITY_VERIFY_ROOTHASH_SIG=y
CONFIG_DM_VERITY_VERIFY_ROOTHASH_SIG_FORCE=y
Then test the negative path: an image whose root hash is unsigned, or signed with a key outside the trusted keyrings, must fail to activate. A verified-boot feature that has never been seen rejecting anything is not yet verified boot.
32-bit Arm supersection mappings removed
Supersections were 16 MiB mappings on ARMv6-class and XScale3 MMUs that carried extra physical address bits. They were the only way a non-LPAE 32-bit Arm kernel could reach physical addresses above 4 GiB, giving 36-bit addressing on hardware that predates LPAE. Linux 7.4 removes them, in a change by Karl Mehltretter merged through the Arm tree.
The consequences are precise. MAX_PHYSMEM_BITS for non-LPAE kernels drops from 36 to 32. CONFIG_IO_36 and the 36-bit ioremap path are gone, and a non-LPAE ioremap() now rejects ranges extending past 4 GiB. The ioremap alignment requirement falls from 16 MiB to 2 MiB as a side effect. EFI runtime regions above 4 GiB are skipped on non-LPAE kernels, which disables runtime services but lets the system boot. LPAE kernels are untouched.
The change started as a build-failure fix — the 36-bit default overflowed a 32-bit resource size in some configurations — and widened when the analysis showed no remaining user: the last XScale3 platform in the tree, PXA3xx, has nothing mapped that high. It fits the broader trimming of 32-bit Arm that we covered in the 7.3 deprecations and, this week, in the proposed memory-split default changes — a separate series, still under review, from the same cleanup effort.
How to check your device
Boards running LPAE kernels are unaffected, and so is every non-LPAE board whose RAM and device registers all sit below 4 GiB, which is nearly all of them. The check is one line:
raghu@techveda.org:~$ zcat /proc/config.gz | grep ARM_LPAE
# CONFIG_ARM_LPAE is not set
A non-LPAE design that genuinely used 36-bit physical addressing has two options going forward: enable LPAE where the core supports it, or stay on a kernel series older than 7.4. If a device-tree node on your board declares memory or registers above 0xffffffff without LPAE, treat that as a finding, not a detail.
Raspberry Pi 5 RTC and IOMMU drivers reach mainline
Two more pieces of Raspberry Pi 5 hardware support landed in Linux 7.4. A dedicated driver for the board’s real-time clock, rtc-raspberrypi, was contributed by Sander Speetjens, and Daniel Drake added a driver for the IOMMUs in the BCM2712 SoC, including their cache handling. Neither exists in 7.3; both are new files in the 7.4 tree.
The Pi 5 has been usable with mainline kernels for some time, but with gaps that kept serious products on Raspberry Pi’s downstream tree. Each of these drivers narrows that gap: the RTC matters for timekeeping and wake behaviour in disconnected deployments, and the IOMMU driver is groundwork for safe DMA and for the peripherals that sit behind those IOMMUs. The same SoC powers the Compute Module 5, so the industrial side of the Pi ecosystem benefits equally.
There is a scheduling catch. Linux 7.3 is the 2026 LTS, and these drivers are not in it. A product that pins 7.3.y for its six years of stable updates will not receive them through that channel; mainline Pi 5 support of this generation starts at 7.4 and will reach an LTS only with the 2027 selection.
What it means for Pi 5-based products
Teams planning to run distro or mainline kernels on Pi 5 or CM5 hardware should put a 7.4 release candidate on a bench board now and record what still falls back to the downstream tree in their configuration. The downstream kernel remains the complete option today. For products choosing between the 7.3 LTS and tracking newer stable series, these drivers belong on the comparison sheet.
Renesas RZ/T2N pairs Linux with four-axis motor control
Renesas announced the RZ/T2N on 5 November, and the silicon is orderable now together with its evaluation kit. The part combines two Cortex-A55 cores at up to 1.2 GHz for application processing with two Cortex-R52 cores at up to 800 MHz for real-time control, each R52 carrying 576 KB of tightly coupled memory with ECC. It slots between the existing RZ/T2M, which offers only R52 cores, and the RZ/T2H, which pairs four A55 cores with the same real-time side.
The integration story is the point. One chip controls up to four motor axes with three-phase PWM and eight encoder interface channels speaking the common industrial encoder protocols, while networking comes from three Gigabit Ethernet MACs, a three-port TSN switch and, on selected variants, a hardware EtherCAT SubDevice Controller, with PROFINET RT/IRT and EtherNet/IP also supported. External memory is 16-bit LPDDR4-1600, up to 2 GB with inline ECC. TrustZone and secure boot cover the security baseline.
On the software side, Renesas ships a Linux BSP based on kernel 6.12 — a longterm-maintained series — with Yocto 5.0, alongside its Flexible Software Package for the R52 side and OpenAMP for communication between the two worlds. The functional safety software package is planned for the second half of 2027. Distributor listings put the MPU near $50 and the evaluation kit at $583.
How to evaluate it
Treat the RZ/T2N as a consolidation candidate: designs that today pair an application processor with a separate motor-control MCU and an EtherCAT ASIC can be collapsed onto one device. Measure the Linux side against your HMI and connectivity load first, since two A55 cores is deliberately modest. Check which variant carries the EtherCAT controller and the crypto features your design needs before committing, and if functional safety certification is on your roadmap, plan around the 2027 software date rather than the silicon availability.
sbom-cve-check 1.4.0 adds OpenVEX export
Bootlin released version 1.4.0 of sbom-cve-check on 10 November. The tool, developed by Benjamin Robin, is a standalone checker that takes an SPDX 2.2 or 3.0 software bill of materials and matches it against the NVD and CVE List databases, reporting which components carry known vulnerabilities. It can already write results as an SPDX 3.0 document or in the Yocto Project cve-check manifest format, which makes it a drop-in companion for the SBOM workflow that Yocto 6.1 Blacksail sharpened last month.
Two additions in 1.4.0 are worth the upgrade. The new OpenVEX exporter produces machine-readable exploitability statements — the format in which a vendor records that a reported CVE does or does not affect a product and why. Customers and regulators increasingly expect such statements to accompany an SBOM rather than a spreadsheet of triage notes. The release also substantially improves the analysis of Linux kernel CVE version ranges. The kernel CVE authority publishes explicit vulnerable-and-fixed version pairs per CVE, at a volume of thousands per year, and naive version comparison against that data routinely misclassifies point releases; better range analysis directly reduces false findings on the component that dominates most embedded CVE reports.
How to use it
Generate an SPDX SBOM from your Yocto build with the create-spdx class, install sbom-cve-check from PyPI, and run it against that document. Compare its kernel findings with your current cve-check output for one release before switching tools, so the differences are understood. Then start publishing OpenVEX statements for the CVEs your triage closes.
References
- MAINTAINERS: remove my reviewerships, add CREDITS entry for THP (lore.kernel.org thread)
- dm-verity: add DM_VERITY_VERIFY_ROOTHASH_SIG_FORCE (commit)
- ARM: 9486/1: set !LPAE MAX_PHYSMEM_BITS to 32 and remove supersection mappings (commit)
- rtc: Add Raspberry Pi 5 RTC driver (commit)
- iommu: Add Broadcom BCM2712 IOMMU driver (commit)
- The first part of the 7.4 merge window
- The second part of the 7.4 merge window
- Renesas RZ/T2N announcement
- RZ/T2N specifications and pricing
- sbom-cve-check 1.4.0 released
- kernel.org releases
— Raghu Bharadwaj



