Linux developer Cong Wang has unveiled Kmorph, an experimental tool that keeps an already-booted second kernel on standby so it can take over a machine's CPUs, memory and devices after a crash in about a second — without a reboot. Built on the multikernel extensions for Linux, the same mechanism can also apply planned kernel upgrades with no downtime.
The foundation is the multikernel architecture: several Linux kernels run in parallel on the same machine, each owning its own CPUs, memory regions and devices, with no hypervisor needed. Kmorph boots the standby kernel in advance on a few configurable CPUs and part of the RAM, where a daemon called kmorphd watches the incumbent kernel over vsock.
When replies stop coming, the takeover begins. The successor first stops the old kernel's CPUs and confirms they have actually halted — fencing that prevents both kernels from touching the same resources. If it cannot safely stop a CPU, it does not proceed. It then takes over the devices, resets them where possible, restores the network configuration, and brings the remaining CPUs and available memory back in via hotplug. Wang puts the gap from fault detection to running successor at roughly one second; a prepared Nginx service served requests within two seconds of the kernel crash in tests.
On request, Kmorph also captures a kdump-compatible vmcore with the crashed kernel's memory contents, CPU registers and VMCOREINFO. Unlike a conventional kdump flow, no capture kernel needs to start after the crash and the machine does not have to boot back into the production system.
No reboot does not mean no interruption: Kmorph does not resume the old kernel's running processes, so userspace normally restarts after a takeover. For selected services it can pre-start standby instances inside the reserve kernel — for example with the configuration standby = nginx.service — which are then carried across with a pivot_root onto the regular root filesystem. Services with their own mount namespaces, such as those using PrivateTmp= or ProtectSystem=, must instead be restarted, losing the benefit of the early start.
The same takeover can be triggered deliberately with kmorph upgrade while the incumbent kernel still works. The new kernel can then arm a further successor of its own; distribution integration leans on dracut or initramfs-tools, kernel-install hooks and a systemd service. x86-64 and ARM64 are supported today.
An additional practical test comes from Clement Verna: a Fedora 45 bootc integration with a multikernel kernel and Kmorph, exercised in a KVM VM with two vCPUs and 16 GB of RAM on a Graviton3 host, with SELinux disabled. In that build the successor cannot yet start a further reserve kernel, so only one switch per boot is possible. Verna explicitly calls the setup a proof of concept and warns against production use.



