Multikernel Linux v7.0-mk5 Brings Bare-Metal Split-Kernel Isolation to ARM64 With IOMMU Device Leases
Multikernel Linux (`mklinux`) has officially broken out of its x86_64-only roots. On November 17–18, 2026, maintainer Cong Wang released `v7.0-mk5`—an out-of-tree 114-patch series built atop the upstream Linux 7.0 kernel—bringing full bare-metal multikernel support and prebuilt Debian/Ubuntu and Fedora packages to 64-bit ARM (`arm64` / `aarch64`) hardware for the first time. Alongside ARM64 bring-up, the release introduces strict IOMMU confinement for lent PCI devices, Intel VT-d interrupt remapping delegation, chained rebootless kernel takeovers, and an important `kexec_file_load()` ABI break that requires rebuilding user-space loaders. ### How Multikernel Isolates Cores Without a Hypervisor In a conventional Linux kernel, user-space workloads, device drivers, and hardware/software interrupt handlers all share the same kernel image and can contend for the same CPU cores. Even aggressive kernel tuning with `isolcpus`, `irqaffinity`, or threaded NAPI cannot structurally prevent bottom-half work from spilling onto application cores, while hypervisors like KVM introduce virtualization overhead and trap-and-emulate costs. Multikernel takes a split-kernel approach directly on bare metal. As LWN's technical overview and Multikernel Technologies' architecture documentation explain, a primary host kernel (instance 0) owns the initial pool of CPUs, physical memory, and PCI devices, then carves subsets out for independent "spawn" kernels booted natively via `kexec_file_load()`. Nothing is emulated or trapped: each spawn kernel runs on its own physical cores and RAM, communicating across isolated lock-free shared-memory rings while offloading general I/O or keeping dedicated real-time or crash-recovery workloads isolated from driver faults. ### Cracking ARM64 Boot, GICv3 Tenancy, and Force Halts Porting that model from x86_64 to ARM64 required solving hardware handoff challenges specific to ARM's interrupt and firmware stack. Rather than giving each spawn kernel full ownership of an ARM system's Generic Interrupt Controller (GICv3), `v7.0-mk5` treats spawns as GICv3 tenants: the host initializes the shared distributor and LPI tables, while each spawn kernel initializes only the redistributors and CPU interfaces belonging to its own assigned cores. CPUs are brought up, verified offline, and released through ARM's Power State Coordination Interface (PSCI), booting a standard `arm64` `Image` from a host-constructed device tree with SGIs sent by MPIDR acting as doorbells. Crucially for crash isolation, if a spawn kernel locks up or crashes, the host can force-halt its cores using ARM Software Delegated Exception Interface (SDEI) event 0—reaching a wedged CPU even when every normal exception is masked—and per-parent/child IPI message rings ensure one dying spawn can no longer wedge communications for sibling instances. Spawn instances on ARM64 remain restricted to PCI devices (routing MSIs through the host's Interrupt Translation Service), excluding non-PCI platform devices. ### Device Leases, IOMMU Confinement, and Chained Takeovers Two major systems-engineering upgrades in `v7.0-mk5` address safety and continuity across crashes on both x86_64 and ARM64: - **IOMMU-confined device leases:** When a host lends a PCI function to a spawn kernel, a `mk_pciback` stub driver locks the host out from rebinding it, and the IOMMU restricts that device's DMA strictly to the spawn's granted memory pages. When the lease ends, the function is reset while DMA confinement remains active. - **Chained takeovers and x86 park tables:** Previously, handing machine control from a failing host kernel to a successor (used in rebootless crash-recovery setups such as Kmorph) had single-hop limits. In `v7.0-mk5`, a successor can fence its host one CPU at a time, claim host VT-d units or ARM64 ITSes without disabling address translation, adopt normal memory as kernel memory, and arm a subsequent successor indefinitely—backed on x86 by a new per-CPU park table. ### Breaking ABI and Real-World Packaging Caveats Developers experimenting with `mklinux` cannot drop `v7.0-mk5` in blindly. The release moves the multikernel flag and instance ID to the upper half of `kexec_file_load()` flags (`KEXEC_FILE_MULTIKERNEL` and `KEXEC_FILE_MK_ID_MASK`) to eliminate a collision with `KEXEC_FILE_NO_CMA` and `KEXEC_FILE_FORCE_DTB`, breaking binary compatibility with loaders compiled against `v7.0-mk4` or earlier. Prebuilt `.deb` packages (built against Ubuntu 26.04 generic configs with `CONFIG_MULTIKERNEL=y`) and `.rpm` packages for `amd64` and `arm64` omit debug info/BTF and are not Secure Boot signed. Yet as heterogeneous ARM servers, high-core-count workstations, and isolated real-time partitions proliferate, `v7.0-mk5` marks the clearest milestone yet toward running multiple independent Linux kernels side by side on modern ARM silicon. ### Sources & Further Reading - [Phoronix (Release announcement and kernel context)](https://www.phoronix.com/news/Multi-Kernel-mklinux-7.0-mk5) - [GitHub multikernel/linux (Primary v7.0-mk5 release notes and changelog)](https://github.com/multikernel/linux/releases/tag/v7.0-mk5) - [Multikernel Technologies (Split-kernel architecture documentation)](https://multikernel.io/technology.html) - [LWN.net (Technical background on mklinux design)](https://lwn.net/Articles/1090582/) ~Linux ~ARM64 ~Kernel ~Multikernel