Kexec reboot consistently fails on ARM64 Hyper-V guests (Azure VMs). During the kexec shutdown path, VMBus channels are never cleaned up, so in the fresh kexec kernel startup hv_acpi_init() blocks because the hypervisor still holds the old kernel's VMBus session open.
This is because ARM64 lacks the VMBus teardown that x86 performs during kexec via hv_machine_shutdown(). On x86, machine_ops.shutdown is overridden to send CHANNELMSG_UNLOAD and disable SynIC before CPUs go offline. ARM64 has no equivalent mechanism. Why existing notification mechanisms don't work: - Reboot notifiers: Fire too early - they run inside kernel_restart_prepare(), BEFORE device_shutdown(). VMBus UNLOAD must happen AFTER device_shutdown() so that PCI drivers can clean up interrupt mappings (PCI_DELETE_INTERRUPT_MESSAGE) before UNLOAD force-closes channels. Sending UNLOAD first causes interrupt mapping leaks because the mappings aren't released on channel close. - Device .shutdown callbacks: Too narrow - these handle per-device cleanup (e.g., individual VMBus channel teardown), but cannot perform bus-level operations like sending the global VMBus UNLOAD message or removing the SynIC CPU hotplug state via cpuhp_remove_state(). - VMBus parent device .shutdown callback: Also insufficient - cpuhp_remove_state() for SynIC teardown must run after all device shutdown completes, not during it. A parent .shutdown callback still executes within device_shutdown(). - The required window is: after device_shutdown() completes, after cpu_hotplug_enable(), but before smp_shutdown_nonboot_cpus(). That window exists inside machine_shutdown(), which currently has no hook. The need for an ARM64 shutdown hook was previously discussed in [1] but lacked a concrete failure case at the time. We now have one. [1] https://lore.kernel.org/linux-arm-kernel/[email protected]/ This RFC proposes fixing kexec on ARM64 Hyper-V guests with: Patch 1: A platform hook (arm64_pre_smp_shutdown_hook) in ARM64's machine_shutdown(), analogous to x86's machine_ops.shutdown. This runs after device_shutdown() and cpu_hotplug_enable(), allowing platform code to inject pre-shutdown logic at the right point in the kexec path. Design choices I'd like feedback on: - Single function pointer vs full machine_ops struct: ARM64 uses kernel-wide APIs (register_restart_handler, register_platform_power_off) for restart/poweroff rather than x86's monolithic machine_ops, so a targeted hook seemed more consistent with the ARM64 pattern. - On ARM64, machine_shutdown() is only called from kernel_kexec(), unlike x86/powerpc where it's also called from restart/halt/poweroff. Patch 2: Uses this hook to call hv_kexec_handler() which performs: - vmbus_initiate_unload(): sends CHANNELMSG_UNLOAD to host - cpuhp_remove_state(): disables SynIC (SIMP, SIEFP, SINT) on all CPUs, ensuring the kexec'd kernel starts with clean state Tested on ARM64 Azure VMs (Ubuntu 22.04, multiple vCPU configs): - kexec reboot succeeds, VM comes back online with SSH - Normal reboot/poweroff unaffected - Multiple consecutive kexec cycles pass Looking for feedback on: 1. Is the single function pointer hook acceptable for ARM64, or would the ARM64 maintainers prefer a registration API or __weak function? 2. Should this be arm64-specific or generic kexec infrastructure? Shradha Gupta (2): arm64: Add pre-shutdown hook to machine_shutdown() arm64/hyperv: Add kexec handler using machine_shutdown hook arch/arm64/hyperv/mshyperv.c | 37 ++++++++++++++++++++++++++++ arch/arm64/include/asm/system_misc.h | 2 ++ arch/arm64/kernel/process.c | 13 ++++++++++ 3 files changed, 52 insertions(+) -- 2.43.0
