Re: [Qemu-devel] [RFC PATCH v3 12/24] spapr: CPU hotplug support

[Bharata B Rao]

On Tue, May 05, 2015 at 04:59:51PM +1000, David Gibson wrote:
> On Fri, Apr 24, 2015 at 12:17:34PM +0530, Bharata B Rao wrote:
> > Support CPU hotplug via device-add command. Set up device tree
> > entries for the hotplugged CPU core and use the exising EPOW event
> > infrastructure to send CPU hotplug notification to the guest.
> > 
> > Also support cold plugged CPUs that are specified by -device option
> > on cmdline.
> > 
> > Signed-off-by: Bharata B Rao <address@hidden>
> > ---
> >  hw/ppc/spapr.c        | 129 
> > ++++++++++++++++++++++++++++++++++++++++++++++++++
> >  hw/ppc/spapr_events.c |   8 ++--
> >  hw/ppc/spapr_rtas.c   |  11 +++++
> >  3 files changed, 145 insertions(+), 3 deletions(-)
> > 
> > diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> > index b526b7d..9b0701c 100644
> > --- a/hw/ppc/spapr.c
> > +++ b/hw/ppc/spapr.c
> > @@ -33,6 +33,7 @@
> >  #include "sysemu/block-backend.h"
> >  #include "sysemu/cpus.h"
> >  #include "sysemu/kvm.h"
> > +#include "sysemu/device_tree.h"
> >  #include "kvm_ppc.h"
> >  #include "mmu-hash64.h"
> >  #include "qom/cpu.h"
> > @@ -662,6 +663,17 @@ static void spapr_populate_cpu_dt(CPUState *cs, void 
> > *fdt, int offset)
> >      unsigned sockets = opts ? qemu_opt_get_number(opts, "sockets", 0) : 0;
> >      uint32_t cpus_per_socket = sockets ? (smp_cpus / sockets) : 1;
> >      uint32_t pft_size_prop[] = {0, cpu_to_be32(spapr->htab_shift)};
> > +    sPAPRDRConnector *drc;
> > +    sPAPRDRConnectorClass *drck;
> > +    int drc_index;
> > +
> > +    if (spapr->dr_cpu_enabled) {
> > +        drc = spapr_dr_connector_by_id(SPAPR_DR_CONNECTOR_TYPE_CPU, index);
> > +        g_assert(drc);
> > +        drck = SPAPR_DR_CONNECTOR_GET_CLASS(drc);
> > +        drc_index = drck->get_index(drc);
> > +        _FDT((fdt_setprop_cell(fdt, offset, "ibm,my-drc-index", 
> > drc_index)));
> > +    }
> >  
> >      _FDT((fdt_setprop_cell(fdt, offset, "reg", index)));
> >      _FDT((fdt_setprop_string(fdt, offset, "device_type", "cpu")));
> > @@ -1850,6 +1862,114 @@ static void spapr_nmi(NMIState *n, int cpu_index, 
> > Error **errp)
> >      }
> >  }
> >  
> > +static void *spapr_populate_hotplug_cpu_dt(DeviceState *dev, CPUState *cs,
> > +                                            int *fdt_offset)
> > +{
> > +    PowerPCCPU *cpu = POWERPC_CPU(cs);
> > +    DeviceClass *dc = DEVICE_GET_CLASS(cs);
> > +    int id = ppc_get_vcpu_dt_id(cpu);
> > +    void *fdt;
> > +    int offset, fdt_size;
> > +    char *nodename;
> > +
> > +    fdt = create_device_tree(&fdt_size);
> > +    nodename = g_strdup_printf("address@hidden", dc->fw_name, id);
> > +    offset = fdt_add_subnode(fdt, 0, nodename);
> > +
> > +    spapr_populate_cpu_dt(cs, fdt, offset);
> > +    g_free(nodename);
> > +
> > +    *fdt_offset = offset;
> > +    return fdt;
> > +}
> > +
> > +static void spapr_cpu_plug(HotplugHandler *hotplug_dev, DeviceState *dev,
> > +                            Error **errp)
> > +{
> > +    CPUState *cs = CPU(dev);
> > +    PowerPCCPU *cpu = POWERPC_CPU(cs);
> > +    int id = ppc_get_vcpu_dt_id(cpu);
> > +    sPAPRDRConnector *drc =
> > +        spapr_dr_connector_by_id(SPAPR_DR_CONNECTOR_TYPE_CPU, id);
> > +    sPAPRDRConnectorClass *drck;
> > +    int smt = kvmppc_smt_threads();
> > +    Error *local_err = NULL;
> > +    void *fdt = NULL;
> > +    int i, fdt_offset = 0;
> > +
> > +    /* Set NUMA node for the added CPUs  */
> > +    for (i = 0; i < nb_numa_nodes; i++) {
> > +        if (test_bit(cs->cpu_index, numa_info[i].node_cpu)) {
> > +            cs->numa_node = i;
> > +            break;
> > +        }
> > +    }
> > +
> > +    /*
> > +     * SMT threads return from here, only main thread (core) will
> > +     * continue and signal hotplug event to the guest.
> > +     */
> > +    if ((id % smt) != 0) {
> > +        return;
> > +    }
> 
> Couldn't you avoid this by attaching this call to the core device,
> rather than the individual vcpu thread objects?

Adding a socket device will result in realize call for that device.
Socket realizefn will call core realizefn and core realizefn will call
thread (or CPU) realizefn. device_set_realized() will call ->plug handler
for all these devices (socket, cores and threads) and that's how we
end up here even for threads.

This will be same when I get rid of socket abstraction and hot plug
cores for the same reason as above.

And calling ->plug handler in the context of threads is required
to initialize board specific CPU bits for the CPU thread that is being
realized.

> 
> 
> > +    if (!spapr->dr_cpu_enabled) {
> > +        /*
> > +         * This is a cold plugged CPU but the machine doesn't support
> > +         * DR. So skip the hotplug path ensuring that the CPU is brought
> > +         * up online with out an associated DR connector.
> > +         */
> > +        return;
> > +    }
> > +
> > +    g_assert(drc);
> > +
> > +    /*
> > +     * Setup CPU DT entries only for hotplugged CPUs. For boot time or
> > +     * coldplugged CPUs DT entries are setup in spapr_finalize_fdt().
> > +     */
> > +    if (dev->hotplugged) {
> > +        fdt = spapr_populate_hotplug_cpu_dt(dev, cs, &fdt_offset);
> > +    }
> > +
> > +    drck = SPAPR_DR_CONNECTOR_GET_CLASS(drc);
> > +    drck->attach(drc, dev, fdt, fdt_offset, !dev->hotplugged, &local_err);
> > +    if (local_err) {
> > +        g_free(fdt);
> > +        error_propagate(errp, local_err);
> > +        return;
> > +    }
> > +
> > +    /*
> > +     * We send hotplug notification interrupt to the guest only in case
> > +     * of hotplugged CPUs.
> > +     */
> > +    if (dev->hotplugged) {
> > +        spapr_hotplug_req_add_event(drc);
> > +    } else {
> > +        /*
> > +         * HACK to support removal of hotplugged CPU after VM migration:
> > +         *
> > +         * Since we want to be able to hot-remove those coldplugged CPUs
> > +         * started at boot time using -device option at the target VM, we 
> > set
> > +         * the right allocation_state and isolation_state for them, which 
> > for
> > +         * the hotplugged CPUs would be set via RTAS calls done from the
> > +         * guest during hotplug.
> > +         *
> > +         * This allows the coldplugged CPUs started using -device option to
> > +         * have the right isolation and allocation states as expected by 
> > the
> > +         * CPU hot removal code.
> > +         *
> > +         * This hack will be removed once we have DRC states migrated as 
> > part
> > +         * of VM migration.
> > +         */
> > +        drck->set_allocation_state(drc, SPAPR_DR_ALLOCATION_STATE_USABLE);
> > +        drck->set_isolation_state(drc, 
> > SPAPR_DR_ISOLATION_STATE_UNISOLATED);
> > +    }
> > +
> > +    return;
> > +}
> > +
> >  static void spapr_machine_device_plug(HotplugHandler *hotplug_dev,
> >                                        DeviceState *dev, Error **errp)
> >  {
> > @@ -1858,6 +1978,15 @@ static void spapr_machine_device_plug(HotplugHandler 
> > *hotplug_dev,
> >          PowerPCCPU *cpu = POWERPC_CPU(cs);
> >  
> >          spapr_cpu_init(cpu);
> > +        spapr_cpu_reset(cpu);
> 
> I'm a little surprised these get called here, rather than in the
> creation / realize path of the core qdev.

These two routines (spapr_cpu_init and spapr_cpu_reset) are sPAPR or
board specific and I can't even have them as part of ppc_cpu_realizefn.

We had discussed this earlier at
http://lists.gnu.org/archive/html/qemu-devel/2015-02/msg04399.html

Regards,
Bharata.