On Thu, 9 Oct 2025 12:32:15 +0000 Miguel Luis <miguel.luis@oracle.com> wrote:

Hi Salil,

...

On 1 Oct 2025, at 01:01, salil.mehta@opnsrc.net wrote:

From: Salil Mehta <salil.mehta@huawei.com>

To support a vCPU hot-add–like model on ARM, the virt machine may be setup with more CPUs than are active at boot. These additional CPUs are fully realized in KVM and listed in ACPI tables from the start, but begin in a disabled state. They can later be brought online or taken offline under host or platform policy control. The CPU topology is fixed at VM creation time and cannot change dynamically on ARM. Therefore, we must determine precisely the 'maxcpus' value that applies for the full lifetime of the VM.

On ARM, this deferred online-capable model is only valid if: - The GIC version is 3 or higher, and - Each non-boot CPU’s GIC CPU Interface is marked “online-capable” in its ACPI GICC structure (UEFI ACPI Specification 6.5, §5.2.12.14, Table 5.37 “GICC CPU Interface Flags”), and - The chosen accelerator supports safe deferred CPU online: * TCG with multi-threaded TCG (MTTCG) enabled * KVM (on supported hosts) * Not HVF or QTest

This patch sizes the machine’s max-possible CPUs during VM init: - If all conditions are satisfied, retain the full set of CPUs corresponding to (`-smp cpus` + `-smp disabledcpus`), allowing the additional (initially disabled) CPUs to participate in later policy-driven online. - Otherwise, clamp the max-possible CPUs to the boot-enabled count (`-smp disabledcpus=0` equivalent) to avoid advertising CPUs the guest can never use.

A new MachineClass flag, `has_online_capable_cpus`, records whether the machine supports deferred vCPU online. This is usable by other machine types as well.

By the definition of

* @has_hotpluggable_cpus: * If true, board supports CPUs creation with -device/device_add.

in include/hw/boards.h

It should be fine to rename it to has_pluggable_cpus. But we should add support to arm/virt for -device/device_add cpu_foo, to avoid awkward -device-set and mangling of -smp. device_add in arm/virt case probably should be limited non hotplug usecase.

seems one could take advantage of MachineClass's has_hotpluggable_cpus variable instead of creating a new has_online_capable_cpus one. (Again, IMHO ‘online capable’ is ACPI nomenclature and doesn’t need to be brought in MachineClass’s)

the issue with has_hotpluggable_cpus might be QMP ABI, where libvirt migh use it to figure out if certain command are supported. CCing libvirt to check if that would break something.

Variable which would be initialized in machvirt_init on an assignment based on GIC version and/or wether there's inactive CPUs and proceed from there anyways, making the default assignment in machine_virt_class_init superfluous.

We're at hw/arm/virt and we know these CPUs are administratively power state coordinated so admin_power_state_supported can still be set there in the presence of inactive CPUs.

Thanks Miguel

...

Signed-off-by: Salil Mehta <salil.mehta@huawei.com> --- hw/arm/virt.c | 84 ++++++++++++++++++++++++++++++--------------- include/hw/boards.h | 1 + 2 files changed, 57 insertions(+), 28 deletions(-)

diff --git a/hw/arm/virt.c b/hw/arm/virt.c index ef6be3660f..76f21bd56a 100644 --- a/hw/arm/virt.c +++ b/hw/arm/virt.c @@ -2168,8 +2168,7 @@ static void machvirt_init(MachineState *machine) bool has_ged = !vmc->no_ged; unsigned int smp_cpus = machine->smp.cpus; unsigned int max_cpus = machine->smp.max_cpus; - - possible_cpus = mc->possible_cpu_arch_ids(machine); + DeviceClass *dc;

/* * In accelerated mode, the memory map is computed earlier in kvm_type() @@ -2186,7 +2185,7 @@ static void machvirt_init(MachineState *machine) * we are about to deal with. Once this is done, get rid of * the object. */ - cpuobj = object_new(possible_cpus->cpus[0].type); + cpuobj = object_new(machine->cpu_type); armcpu = ARM_CPU(cpuobj);

pa_bits = arm_pamax(armcpu); @@ -2201,6 +2200,57 @@ static void machvirt_init(MachineState *machine) */ finalize_gic_version(vms);

+ /* + * The maximum number of CPUs depends on the GIC version, or on how + * many redistributors we can fit into the memory map (which in turn + * depends on whether this is a GICv3 or v4). + */ + if (vms->gic_version == VIRT_GIC_VERSION_2) { + virt_max_cpus = GIC_NCPU; + } else { + virt_max_cpus = virt_redist_capacity(vms, VIRT_GIC_REDIST); + if (vms->highmem_redists) { + virt_max_cpus += virt_redist_capacity(vms, VIRT_HIGH_GIC_REDIST2); + } + } + + if ((tcg_enabled() && !qemu_tcg_mttcg_enabled()) || hvf_enabled() || + qtest_enabled() || vms->gic_version == VIRT_GIC_VERSION_2) { + max_cpus = machine->smp.max_cpus = smp_cpus; + if (mc->has_online_capable_cpus) { + if (vms->gic_version == VIRT_GIC_VERSION_2) { + warn_report("GICv2 does not support online-capable CPUs"); + } + mc->has_online_capable_cpus = false; + } + } + + if (mc->has_online_capable_cpus) { + max_cpus = smp_cpus + machine->smp.disabledcpus; + machine->smp.max_cpus = max_cpus; + } + + if (max_cpus > virt_max_cpus) { + error_report("Number of SMP CPUs requested (%d) exceeds max CPUs " + "supported by machine 'mach-virt' (%d)", + max_cpus, virt_max_cpus); + if (vms->gic_version != VIRT_GIC_VERSION_2 && !vms->highmem_redists) { + error_printf("Try 'highmem-redists=on' for more CPUs\n"); + } + + exit(1); + } + + dc = DEVICE_CLASS(object_class_by_name(machine->cpu_type)); + if (!dc) { + error_report("CPU type '%s' not registered", machine->cpu_type); + exit(1); + } + dc->admin_power_state_supported = mc->has_online_capable_cpus; + + /* uses smp.max_cpus to initialize all possible vCPUs */ + possible_cpus = mc->possible_cpu_arch_ids(machine); + if (vms->secure) { /* * The Secure view of the world is the same as the NonSecure, @@ -2235,31 +2285,6 @@ static void machvirt_init(MachineState *machine) vms->psci_conduit = QEMU_PSCI_CONDUIT_HVC; }

- /* - * The maximum number of CPUs depends on the GIC version, or on how - * many redistributors we can fit into the memory map (which in turn - * depends on whether this is a GICv3 or v4). - */ - if (vms->gic_version == VIRT_GIC_VERSION_2) { - virt_max_cpus = GIC_NCPU; - } else { - virt_max_cpus = virt_redist_capacity(vms, VIRT_GIC_REDIST); - if (vms->highmem_redists) { - virt_max_cpus += virt_redist_capacity(vms, VIRT_HIGH_GIC_REDIST2); - } - } - - if (max_cpus > virt_max_cpus) { - error_report("Number of SMP CPUs requested (%d) exceeds max CPUs " - "supported by machine 'mach-virt' (%d)", - max_cpus, virt_max_cpus); - if (vms->gic_version != VIRT_GIC_VERSION_2 && !vms->highmem_redists) { - error_printf("Try 'highmem-redists=on' for more CPUs\n"); - } - - exit(1); - } - if (vms->secure && !tcg_enabled() && !qtest_enabled()) { error_report("mach-virt: %s does not support providing " "Security extensions (TrustZone) to the guest CPU", @@ -3245,6 +3270,9 @@ static void virt_machine_class_init(ObjectClass *oc, const void *data) hc->plug = virt_machine_device_plug_cb; hc->unplug_request = virt_machine_device_unplug_request_cb; hc->unplug = virt_machine_device_unplug_cb; + + mc->has_online_capable_cpus = true; + mc->nvdimm_supported = true; mc->smp_props.clusters_supported = true; mc->auto_enable_numa_with_memhp = true; diff --git a/include/hw/boards.h b/include/hw/boards.h index 2b182d7817..b27c2326a2 100644 --- a/include/hw/boards.h +++ b/include/hw/boards.h @@ -302,6 +302,7 @@ struct MachineClass { bool rom_file_has_mr; int minimum_page_bits; bool has_hotpluggable_cpus; + bool has_online_capable_cpus; bool ignore_memory_transaction_failures; int numa_mem_align_shift; const char * const *valid_cpu_types; -- 2.34.1