Re: [Qemu-devel] [RFC PATCH v0 3/6] spapr: Represent boot CPUs as spapr-cpu-core devices

[Bharata B Rao <bharata@linux.vnet.ibm.com>]

On Fri, Feb 26, 2016 at 02:45:35PM +1100, David Gibson wrote:
> On Thu, Feb 25, 2016 at 09:52:39PM +0530, Bharata B Rao wrote:
> > Initialize boot CPUs as spapr-cpu-core devices and create links from
> > machine object to these core devices. These links can be considered
> > as CPU slots in which core devices will get hot-plugged. spapr-cpu-core
> > device's slot property indicates the slot where it is plugged. Information
> > about all the CPU slots can be obtained by walking these links.
> > 
> > Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
> > ---
> >  hw/ppc/spapr.c         | 65 +++++++++++++++++++++++++++++++++++++++++++-------
> >  include/hw/ppc/spapr.h |  3 +++
> >  2 files changed, 60 insertions(+), 8 deletions(-)
> > 
> > diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> > index e214a34..1f0d232 100644
> > --- a/hw/ppc/spapr.c
> > +++ b/hw/ppc/spapr.c
> > @@ -64,6 +64,7 @@
> >  
> >  #include "hw/compat.h"
> >  #include "qemu-common.h"
> > +#include "hw/ppc/spapr_cpu_core.h"
> >  
> >  #include <libfdt.h>
> >  
> > @@ -1720,6 +1721,21 @@ static void spapr_validate_node_memory(MachineState *machine, Error **errp)
> >      }
> >  }
> >  
> > +/*
> > + * Check to see if core is being hot-plugged into an already populated slot.
> > + */
> > +static void spapr_cpu_core_allow_set_link(Object *obj, const char *name,
> > +                                          Object *val, Error **errp)
> > +{
> > +    Object *core = object_property_get_link(qdev_get_machine(), name, NULL);
> > +
> > +    if (core) {
> > +        char *path = object_get_canonical_path(core);
> > +        error_setg(errp, "Slot %s already populated with %s", name, path);
> > +        g_free(path);
> > +    }
> > +}
> > +
> >  /* pSeries LPAR / sPAPR hardware init */
> >  static void ppc_spapr_init(MachineState *machine)
> >  {
> > @@ -1728,7 +1744,6 @@ static void ppc_spapr_init(MachineState *machine)
> >      const char *kernel_filename = machine->kernel_filename;
> >      const char *kernel_cmdline = machine->kernel_cmdline;
> >      const char *initrd_filename = machine->initrd_filename;
> > -    PowerPCCPU *cpu;
> >      PCIHostState *phb;
> >      int i;
> >      MemoryRegion *sysmem = get_system_memory();
> > @@ -1742,6 +1757,8 @@ static void ppc_spapr_init(MachineState *machine)
> >      long load_limit, fw_size;
> >      bool kernel_le = false;
> >      char *filename;
> > +    int spapr_cores = smp_cpus / smp_threads;
> > +    int spapr_max_cores = max_cpus / smp_threads;
> >  
> >      msi_supported = true;
> >  
> > @@ -1800,13 +1817,38 @@ static void ppc_spapr_init(MachineState *machine)
> >      if (machine->cpu_model == NULL) {
> >          machine->cpu_model = kvm_enabled() ? "host" : "POWER7";
> >      }
> > -    for (i = 0; i < smp_cpus; i++) {
> > -        cpu = cpu_ppc_init(machine->cpu_model);
> > -        if (cpu == NULL) {
> > -            error_report("Unable to find PowerPC CPU definition");
> > -            exit(1);
> > +
> > +    spapr->cores = g_malloc0(spapr_max_cores * sizeof(Object));
> > +
> > +    for (i = 0; i < spapr_max_cores; i++) {
> > +        Object *spapr_cpu_core  = object_new(TYPE_SPAPR_CPU_CORE);
> 
> So.. if I understand correctly this will always construct maxcpus
> threads at startup.  I's just that the ones beyond initial cpus won't
> be realized.  Was that your intention?  I thought the plan was to only
> construct the hotplugged cpus when they were hotplugged (in contrast
> to my earlier proposal).

Oops! The intention was to to create only cores for smp_cpus
and let device_add create the rest. Will fix.

> 
> > +        char name[32];
> > +
> > +        object_property_set_str(spapr_cpu_core, machine->cpu_model, "cpu_model",
> > +                                &error_fatal);
> > +        object_property_set_int(spapr_cpu_core, smp_threads, "nr_threads",
> > +                                &error_fatal);
> > +        /*
> > +         * Create links from machine objects to all possible cores.
> > +         */
> > +        snprintf(name, sizeof(name), "%s[%d]", SPAPR_MACHINE_CPU_CORE_PROP, i);
> 
> Why do SPAPR_MACHINE_CPU_CORE_PROP and SPAPR_CPU_CORE_SLOT_PROP get
> #defines, but the other properties get bare strings?  I find the bare
> strings are usually easier to follow.

Let me check the convention regarding properties and stick to the
common usage.

> 
> > +        object_property_add_link(OBJECT(spapr), name, TYPE_SPAPR_CPU_CORE,
> > +                                 (Object **)&spapr->cores[i],
> > +                                 spapr_cpu_core_allow_set_link, 0,
> > +                                 &error_fatal);
> > +
> > +        /*
> > +         * Set the link from machine object to core object for all
> > +         * boot time CPUs specified with -smp and realize them.
> > +         * For rest of the hotpluggable cores this is happens from
> > +         * the core hotplug/realization path.
> > +         */
> > +        if (i < spapr_cores) {
> > +            object_property_set_str(spapr_cpu_core, name,
> > +                                    SPAPR_CPU_CORE_SLOT_PROP, &error_fatal);
> > +            object_property_set_bool(spapr_cpu_core, true, "realized",
> > +                                     &error_fatal);
> >          }
> > -        spapr_cpu_init(spapr, cpu, &error_fatal);
> >      }
> >  
> >      if (kvm_enabled()) {
> > @@ -2209,6 +2251,7 @@ static void spapr_machine_device_plug(HotplugHandler *hotplug_dev,
> >                                        DeviceState *dev, Error **errp)
> >  {
> >      sPAPRMachineClass *smc = SPAPR_MACHINE_GET_CLASS(qdev_get_machine());
> > +    sPAPRMachineState *ms = SPAPR_MACHINE(hotplug_dev);
> >  
> >      if (object_dynamic_cast(OBJECT(dev), TYPE_PC_DIMM)) {
> >          int node;
> > @@ -2245,6 +2288,11 @@ static void spapr_machine_device_plug(HotplugHandler *hotplug_dev,
> >          }
> >  
> >          spapr_memory_plug(hotplug_dev, dev, node, errp);
> > +    } else if (object_dynamic_cast(OBJECT(dev), TYPE_CPU)) {
> > +        CPUState *cs = CPU(dev);
> > +        PowerPCCPU *cpu = POWERPC_CPU(cs);
> > +
> > +        spapr_cpu_init(ms, cpu, errp);
> 
> I tend to thin the spapr_cpu_init() should be done from the core's
> create_threads() function, but I'm willing to be persuaded otherwise.

In addition to core device, every cpu thread device will go through
->plug() invocation via its realization path. And spapr_cpu_init()
is needed for every CPU thread. Let me check if it is possible to
invoke this from core's routines.

Regards,
Bharata.