[ANNOUNCE] mklinux v7.0-mk5: multikernel Linux, now on arm64

From: Cong Wang

Date: Tue Nov 17 2026 - 17:46:14 EST


Hi all,

mklinux v7.0-mk5 is out. mklinux is an out-of-tree branch of Linux that
lets one machine run several Linux kernels side by side without a
hypervisor. A host kernel sets aside a pool of CPUs, memory and PCI
devices, loads another kernel into it with kexec_file_load() and starts
it on those CPUs. Each kernel then runs natively on its own share of the
machine. The kernels talk over shared-memory rings and IPIs, which carry
a console, AF_VSOCK and PCI config space access, and a kernel can take
the whole machine over from its host, which is what kernel upgrade and
crash recovery are built on.

This release is 114 patches on top of v7.0-mk4, and the tree is 330
patches on top of v7.0.

What is new
===========

arm64
Multikernel now runs on arm64 as well as x86_64. CPUs are started,
confirmed off and released through PSCI. An Image is loaded with
kexec_file_load() and boots from a device tree the host builds for
it. A spawned kernel is a tenant of the GICv3: it leaves the
distributor and LPI setup to the host and initialises only the
redistributors and CPU interfaces of its own CPUs. It gets PCI
devices only, with MSIs routed through whichever host ITS serves the
device. A kernel is forced down through SDEI event 0 where firmware
has it, which reaches a CPU that has every exception masked, and
with the stop IPI otherwise. arm64 memory hot-remove learned to
split or refuse a partial removal of a linear map leaf and to free
the page tables and vmemmap it empties.

Device leases and DMA confinement
A PCI function lent to another kernel is held on the host by a stub
driver, as vfio-pci and xen-pciback do, and everything the host sets
up for it hangs off one lease. Its DMA is confined through the IOMMU
to the memory of the kernel that drives it, on arm64 (SMMUv3) and on
x86 (VT-d), and follows that kernel's memory as it grows and
shrinks. The function is reset while still confined when the lease
ends.

VT-d
The host leaves alone the units another kernel drives, lends each
spawned kernel a block of interrupt remapping entries and checks
every entry written into it. A kernel that took the machine over can
claim the units of its fenced host without turning translation off.

Chained takeover
A kernel that took over from its host becomes instance 0, can online
the memory it adopts as kernel memory, and can arm and hand over to
a successor of its own, any number of times. On x86 parked CPUs no
longer wait on slots owned by one kernel: there is one park table
for the machine with a cell per CPU, so any kernel can find and wake
a parked CPU, as PSCI allows on arm64.

Messaging
Each parent/child pair has its own rings, so a kernel that dies in
the middle of a send can only wedge its own link.

Console
A spawned kernel can be handed the host's screen as a
simple-framebuffer, for machines reached only through a BMC's remote
KVM.

ABI change
==========

The multikernel flag and the instance id moved to the upper half of the
kexec_file_load() flags, as KEXEC_FILE_MULTIKERNEL (0x80000000) and
KEXEC_FILE_MK_ID_MASK (0x7fff0000). The old values collided with
KEXEC_FILE_NO_CMA and KEXEC_FILE_FORCE_DTB. A loader built against
v7.0-mk4 or earlier has to be rebuilt; kerf 0.5.0, the user-space tool,
uses the new layout.

Getting it
==========

git: https://github.com/multikernel/linux.git, tag v7.0-mk5
release: https://github.com/multikernel/linux/releases/tag/v7.0-mk5
kerf: https://github.com/multikernel/kerf/releases/tag/v0.5.0
docs: Documentation/multikernel/ in the tree

The release page has the full change list and deb and rpm packages for
x86_64 and arm64, built with the Ubuntu 26.04 generic config plus
CONFIG_MULTIKERNEL=y.

Known limitations
=================

- An arm64 instance gets PCI devices only, no platform devices, since
every SPI goes through the one GIC distributor, which stays with the
host.
- An arm64 spawn has to boot with pseudo-NMIs
(irqchip.gicv3_pseudo_nmi=1) so that the host can stop it when SDEI
is not available.
- The prebuilt packages have no debug info, hence no BTF, and are not
Secure Boot signed.

Testing, bug reports and review are all welcome, on the list or in the
issue tracker at https://github.com/multikernel/linux/issues.

Thanks,
Cong