Skip to main content

TECH VEDA

Corporate on-site training - Submit proposal Pick your modulesSharpen your kernel skills: deep dives, drivers, Yocto, CVEs, careers — updated daily. Read the blog →
Automotive

Real-Time Kernel on a 6.12 Automotive BSP: Where Did the RT Branch Go, and What Does the RT Config Turn Off?

Real-time kernel on a 6.12 automotive BSP: what TI Jacinto, Renesas R-Car and NXP S32 publish, and what TI's RT config fragment changes besides preemption.

Real-Time Kernel on a 6.12 Automotive BSP: Where Did the RT Branch Go, and What Does the RT Config Turn Off?

On 64-bit Arm, the separate real-time branch in a vendor kernel tree is no longer needed from Linux 6.12 onwards, and Texas Instruments has removed it: the Jacinto SDK now builds its RT kernel from the same branch and tag as the standard kernel, using one configuration fragment. That fragment sets CONFIG_PREEMPT_RT and also switches off 81 other options, among them seccomp, the stack protector, address-space randomisation, CPU idle and suspend. Renesas publishes a real-time branch only on its 6.1 kernel for R-Car Gen3, and neither Renesas nor NXP publishes an RT configuration for its 6.12 kernel, although the option can be selected in both trees.

In September this blog read the NXP S32 automotive BSP and found that its move from Linux 6.6 to 6.12 had happened without the real-time branch, which stayed on 6.6. That post said it had not checked any other vendor. This post does. It reads the public kernel trees and Yocto layers for Texas Instruments Jacinto and Renesas R-Car on 20 November 2026, re-reads NXP S32, and asks one question of each: if a team needs a real-time kernel on the vendor’s 6.12 baseline, what does the vendor actually publish? It is for engineers who maintain a vehicle computer’s kernel and must decide whether the vendor’s RT build can ship.

Why the 6.12 boundary changes what a vendor has to carry

Until Linux 6.12, CONFIG_PREEMPT_RT existed in the mainline Kconfig but could not be selected on any architecture. The option depends on ARCH_SUPPORTS_RT, and at v6.11 no architecture selected that symbol. A vendor who wanted a real-time kernel therefore applied the out-of-tree PREEMPT_RT patch series, usually on a second branch that was rebased separately at every stable update.

The change arrived as three small commits, one per architecture. The arm64 one, commit d8fccd9ca5f9, was written by Sebastian Andrzej Siewior on 6 September 2024, committed by Thomas Gleixner on 17 September 2024 and acknowledged by Will Deacon. It adds a single line, select ARCH_SUPPORTS_RT, to arch/arm64/Kconfig; its message says the last known obstacle had been removed by the printk changes that preceded it. The same was done for x86 and RISC-V. I checked the release tags: at v6.11 no architecture Kconfig contains the symbol, and at v6.12, tagged on 17 November 2024, arm64, x86 and RISC-V do.

32-bit Arm was not part of that set. arch/arm/Kconfig does not select ARCH_SUPPORTS_RT at v6.12, v6.18, v6.19 or v7.0; it does at v7.1, tagged on 14 June 2026. The longterm list on kernel.org, read on 20 November 2026, stops at 6.18, so on every series listed there a 32-bit Arm real-time kernel still needs the patch series, and a 64-bit one does not.

The patch series itself still exists. The kernel.org directory for 6.12 held patch-6.12.116-rt22, dated 9 November 2026 and about 18 KB compressed, when I read it. On arm64 it is no longer what makes the option selectable.

Texas Instruments Jacinto: the RT branch became a config fragment

TI documents both sides of this change in its own release notes.

The Processor SDK Linux release notes for J784S4, release 10.01.00, are marked “Released Nov 2024”. They list two kernels. The standard one is on branch ti-linux-6.6.y, tag 10.01.08, kernel version 6.6.44. The real-time one has its own heading, its own branch ti-rt-linux-6.6.y, and its own tag 10.01.08-rt.

The release notes served as the latest version on 20 November 2026 are for release 11.02.02, marked “Released Aug 2026”, covering six Jacinto devices on Yocto Scarthgap. They list one kernel: branch ti-linux-6.12.y, tag 11.02.18, kernel version 6.12.57, head commit dated 13 August 2026. Under that single entry there are two configuration lines. The non-RT kernel is defconfig plus ti_arm64_prune.config. The RT kernel is defconfig plus ti_rt.config plus ti_arm64_prune.config. There is no second branch and no -rt tag.

The Yocto layer says the same thing in code. In meta-ti, the recipe linux-ti-staging-rt_6.12.bb sets its 64-bit branch to ti-linux-6.12.y, the same branch as the standard recipe, and for 64-bit devices differs in appending -rt to the local version and in the fragment list. A commit by Beleswar Padhi of TI dated 11 July 2025 explains the one exception, stating that “all ARM64 devices use the same branch for RT and non-RT builds” while the 32-bit devices get a separate one. The branch list confirms it:

raghu@techveda.org:~$ git ls-remote --heads https://git.ti.com/git/ti-linux-kernel/ti-linux-kernel.git | cut -f2 | grep -E "/ti-(rt-)?linux-6[.](6|12|18)[.]y(-arm32)?$"
refs/heads/ti-linux-6.12.y
refs/heads/ti-linux-6.18.y
refs/heads/ti-linux-6.6.y
refs/heads/ti-rt-linux-6.12.y-arm32
refs/heads/ti-rt-linux-6.18.y-arm32
refs/heads/ti-rt-linux-6.6.y

The 6.6 line has a full real-time branch; the 6.12 and 6.18 lines have one only for 32-bit Arm, which matches the upstream position exactly. ti-linux-6.12.y has no localversion-rt file at its root; ti-rt-linux-6.12.y-arm32 has one containing -rt10, and the 6.18 equivalent contains -rt6. The old ti-rt-linux-6.6.y branch stopped moving on 27 November 2024 at 6.6.58 with -rt45.

For a Jacinto team the standard and real-time kernels can no longer diverge in source: they are one commit. Every difference between them is in a text file of 231 lines, which should be read in full before the RT build is adopted.

What the TI real-time kernel configuration turns off

I read kernel/configs/ti_rt.config at the head of ti-linux-6.12.y. It contains 94 configuration assignments. Seven enable CONFIG_PREEMPT_RT, a 1,000 Hz tick and CONFIG_NO_HZ_FULL with their supporting options, five build crypto algorithms as modules, and one sets the security-module list. The other 81 are set to n. The same file at the head of ti-linux-6.18.y has 91 such lines, having added the nftables options, SCTP and the firmware TPM driver.

The disabled options fall into groups. The security group includes CONFIG_SECCOMP, CONFIG_STACKPROTECTOR, CONFIG_VMAP_STACK, CONFIG_RANDOMIZE_BASE, CONFIG_RANDOMIZE_KSTACK_OFFSET, CONFIG_UNMAP_KERNEL_AT_EL0, CONFIG_SECURITY_SELINUX and CONFIG_SECURITY_APPARMOR. The power-management group is CONFIG_CPU_IDLE, CONFIG_CPU_FREQ, CONFIG_SUSPEND and CONFIG_HOTPLUG_CPU. The observability group includes CONFIG_PERF_EVENTS, CONFIG_KALLSYMS, CONFIG_BPF_SYSCALL and CONFIG_CORESIGHT. A fourth group removes most control-group controllers, including memory, CPU sets and the CPU scheduler controller, and a set of networking options that container runtimes use. Several later Arm architecture features are disabled as well.

The file documents its own reasoning, which is unusual for a vendor configuration. Its comments give measured worst-case latency from cyclictest on AM62x hardware for individual options: 41 microseconds rising to 83 with the later Arm architecture features enabled, 50 rising to 69 with kernel stack offset randomisation, 50 rising to 58 with full read-only protection of the linear map, 50 rising to 70 with kernel page-table isolation, and 50 rising to 90 with the control-group controllers. One block of security options is introduced with a comment that calls the removal a hack and says that most users without aggressive latency goals will want those features enabled.

TI’s fragment is therefore not a description of what PREEMPT_RT requires. It is a configuration tuned for a low worst-case figure on a benchmark, and each line records a trade made for that purpose. Two things follow for a vehicle programme.

First, the measurements in the comments are from AM62x, a different device family from the Jacinto parts the SDK release covers. The fragment is shared across all of TI’s 64-bit devices. I found no Jacinto-specific figures in the file.

Second, a team that takes the RT image as delivered also takes a kernel without seccomp, stack protection, address-space randomisation or suspend. Each may be a requirement from elsewhere in the programme: a container runtime that expects seccomp and control groups, a cybersecurity concept written against ISO/SAE 21434 that assumes kernel hardening, or a unit that must suspend. Because the real-time kernel is now only a fragment, the remedy is small. A team can start from CONFIG_PREEMPT_RT=y alone, add the remaining lines one group at a time, and measure on its own board what each one gains.

One limit: I read the fragment, not a built kernel. Kconfig can override a fragment line when another enabled option selects the same symbol, so the final .config of an RT image may differ in places.

Renesas R-Car: a real-time branch on 6.1, none on 6.12

The public Renesas tree, renesas-rcar/linux-bsp, does not show the TI pattern. It shows the older model still in use on an older kernel, and nothing for real time on the newer one.

Among all the branches in that repository on 20 November 2026, two carry an RT suffix: v6.1.166/rcar-5.3.9-RT.rc1 and v6.1.166/rcar-5.3.10-RT.rc1. The second is also tagged rcar-5.3.10-RT, on a commit dated 29 July 2026. It is Linux 6.1.166 with a localversion-rt file containing -rt61, and its arm64 defconfig sets CONFIG_PREEMPT_RT=y. Its sibling without the suffix, tagged rcar-5.3.10, sets CONFIG_PREEMPT=y and has no such file.

The meta-renesas layer shows who can use it. On the scarthgap branch, last committed on 19 November 2026, the Gen3 kernel recipe selects the RT branch when the machine features include linux-rt, for the machines salvator-x, h3ulcb, m3ulcb, m3nulcb, ebisu, draak and geist. The recipe for the R-Car S4 gateway boards, spider and s4sk, uses the same 6.1.166 kernel and has no RT selection. The recipe for the V3x and V4x boards, which include whitehawk and grayhawk, is on Linux 5.10.235 with CONFIG_PREEMPT=y.

The 6.12 kernel in this tree belongs to the newest family. The layer meta-rcar-gen5, whose recipe describes itself as the kernel for R-Car X5x based boards, moved between two Yocto branches: on scarthgap it points at v6.1.102/rcar-6.0.0.rc12, last committed on 23 April 2026, and on wrynose it points at v6.12.80/rcar-6.1.0, last committed on 29 October 2026. On that 6.12 branch the configuration the recipe builds, renesas_bsp_defconfig, sets CONFIG_PREEMPT=y. There is no localversion-rt, no RT fragment under kernel/configs, and no RT switch in the recipe. The 6.1.102 branch before it also set CONFIG_PREEMPT=y.

For Renesas, then, there is no before-and-after at the 6.12 boundary to report. A published real-time kernel exists for Gen3 only, on 6.1, built the old way. The Gen5 kernel had no public RT variant before 6.12 and has none after. The 6.12 branch has no release tag yet, only a branch name.

NXP S32, re-read on 20 November 2026

The NXP findings from September hold, with newer releases. The kernel repository now has 6.12 branches up to release/s32n7_bsp5.0-6.12.112, last committed on 23 October 2026, and no branch with an -rt suffix is on a kernel newer than 6.6. The most recently tagged release with one is BSP 47.0, release/bsp47.0-6.6.99-rt, with -rt58. On both release/bsp48.0-6.12.92 and the newer S32N7 branch, s32cc_defconfig and s32n7_defconfig set CONFIG_PREEMPT=y, and neither arch/arm64/configs nor kernel/configs contains a vendor RT fragment.

What the TI comparison adds is a correction of emphasis. On arm64 at 6.12 the absence of an -rt branch is expected, because no patch is needed, and the NXP tree selects ARCH_SUPPORTS_RT like every other 6.12 arm64 tree. What NXP does not publish for 6.12 is a real-time kernel configuration, and with it any statement that such a build has been tested. That is a narrower finding than a missing branch.

The three trees side by side

Vendor treePublic 6.12 kernelRT on 6.12Last separate RT branch
TI Jacinto (ti-linux-kernel)6.12.57, SDK 11.02.02, August 2026Same branch, ti_rt.config fragment, documented in the release notesti-rt-linux-6.6.y, 6.6.58, last commit 27 November 2024
Renesas R-Car (linux-bsp)6.12.80, Gen5 branch, no tag, October 2026None published; default is CONFIG_PREEMPTrcar-5.3.10-RT, 6.1.166, 29 July 2026, Gen3 only
NXP S32 (nxp-auto-linux/linux)6.12.112, S32N7 BSP 5.0, October 2026None published; default is CONFIG_PREEMPTbsp47.0-6.6.99-rt, last commit 26 June 2026

How to check the BSP in front of you

Three checks on the host apply to any vendor kernel tree: whether the architecture allows the option, whether the tree still carries the out-of-tree series, and whether the vendor has enabled the option anywhere.

raghu@techveda.org:~$ grep -n "ARCH_SUPPORTS_RT" arch/arm64/Kconfig
raghu@techveda.org:~$ ls localversion-rt
raghu@techveda.org:~$ grep -rl "PREEMPT_RT=y" arch/arm64/configs kernel/configs

The first command prints a line on any tree based on 6.12 or later. The second succeeds only on a tree that still carries the out-of-tree series. An empty result from the third on a 6.12 tree means the vendor ships no real-time configuration, not that one cannot be built.

On the board, the build’s own version string settles which kernel is running:

root@j784s4-evm:~# uname -v

The kernel build places the preemption model in that string, so a real-time kernel reports PREEMPT_RT there and a standard vendor build reports PREEMPT or PREEMPT_DYNAMIC.

What this post does not claim

Everything above is read from public repositories and public release notes. NXP and Renesas both distribute BSP material through channels that need an account, and a real-time kernel configuration for 6.12 may exist there; I can only say that none is in the public trees. I did not build any of these kernels or measure latency on any board: the latency numbers are TI’s, from comments in its fragment, on AM62x. I did not read the Qualcomm or NVIDIA automotive kernels, and the Renesas V4x kernel on 5.10 was checked on one branch only. Nothing here says which preemption model a production vehicle uses, and nothing here is a statement about functional safety: a real-time kernel bounds scheduling latency and makes no claim under ISO 26262.

Key takeaways

  • From Linux 6.12, arm64 selects ARCH_SUPPORTS_RT, so a 64-bit vendor tree needs no patch series and no second branch to build a real-time kernel. 32-bit Arm gained the same only in 7.1.
  • TI Jacinto moved from a separate RT branch and tag in SDK 10.01.00 (November 2024) to one branch, one tag and a fragment in SDK 11.02.02 (August 2026).
  • TI’s ti_rt.config sets 81 options to n besides enabling PREEMPT_RT, including seccomp, the stack protector, KASLR, CPU idle and suspend, with TI’s own AM62x latency measurements as the stated reason.
  • Renesas publishes a real-time kernel only for R-Car Gen3 on 6.1.166. Its 6.12 Gen5 branch and NXP’s 6.12 S32 branches default to CONFIG_PREEMPT and carry no RT configuration.
  • On a 6.12 arm64 BSP a missing -rt branch is normal. Ask whether the vendor publishes an RT configuration, and what else it changes.
Was this worth your time?

Frequently asked questions

Does a 6.12 automotive BSP need the PREEMPT_RT patch to build a real-time kernel?
Not on 64-bit Arm. From v6.12 the arm64 Kconfig selects ARCH_SUPPORTS_RT, so CONFIG_PREEMPT_RT can be enabled in an unpatched tree. 32-bit Arm needs the patch series on every longterm kernel, because mainline support arrived in 7.1.

Where is the TI real-time kernel for Jacinto on 6.12?
On the same branch and tag as the standard kernel. The SDK 11.02.02 release notes give ti-linux-6.12.y, tag 11.02.18, for both, and the RT build adds the ti_rt.config fragment.

Is the TI RT configuration only a change of preemption model?
No. Besides CONFIG_PREEMPT_RT=y the fragment sets 81 options to n, including seccomp, the stack protector, address-space randomisation, CPU idle, CPU frequency scaling and suspend.

Does Renesas publish a real-time kernel for R-Car?
For Gen3 boards, yes: the tag rcar-5.3.10-RT on Linux 6.1.166, selected in meta-renesas by the linux-rt machine feature. The public 6.12 branch for the Gen5 boards has no RT variant.

Further reading

RB
Raghu Bharadwaj

Founder, TECH VEDA — 20+ years teaching the Linux kernel, device drivers and embedded systems.

Follow on LinkedIn

Get new posts by email

Kernel, embedded Linux and AI-era engineering — a few sharp reads a month. No spam.

We email occasionally and never share your address.