Re: [Qemu-ppc] [RFC PATCH v1 08/10] spapr: CPU hotplug support

[Igor Mammedov]

On Wed, 9 Mar 2016 13:58:53 +1100
David Gibson <address@hidden> wrote:

> On Tue, Mar 08, 2016 at 10:37:08AM +0100, Igor Mammedov wrote:
> > On Tue, 8 Mar 2016 15:27:39 +1100
> > David Gibson <address@hidden> wrote:
> >   
> > > On Mon, Mar 07, 2016 at 11:59:42AM +0530, Bharata B Rao wrote:  
> > > > On Mon, Mar 07, 2016 at 02:49:06PM +1100, David Gibson wrote:    
> > > > > On Fri, Mar 04, 2016 at 12:24:19PM +0530, Bharata B Rao wrote:    
> > > > > > 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.
> > > > > > 
> > > > > > Signed-off-by: Bharata B Rao <address@hidden>
> > > > > > ---
> > > > > >  hw/ppc/spapr.c                  | 73 
> > > > > > ++++++++++++++++++++++++++++++++++++++++-
> > > > > >  hw/ppc/spapr_cpu_core.c         | 60 
> > > > > > +++++++++++++++++++++++++++++++++
> > > > > >  hw/ppc/spapr_events.c           |  3 ++
> > > > > >  hw/ppc/spapr_rtas.c             | 24 ++++++++++++++
> > > > > >  include/hw/ppc/spapr.h          |  4 +++
> > > > > >  include/hw/ppc/spapr_cpu_core.h |  2 ++
> > > > > >  6 files changed, 165 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> > > > > > index 5acb612..6c4ac50 100644
> > > > > > --- a/hw/ppc/spapr.c
> > > > > > +++ b/hw/ppc/spapr.c
> > > > > > @@ -603,6 +603,18 @@ static void spapr_populate_cpu_dt(CPUState 
> > > > > > *cs, void *fdt, int offset,
> > > > > >      size_t page_sizes_prop_size;
> > > > > >      uint32_t vcpus_per_socket = smp_threads * smp_cores;
> > > > > >      uint32_t pft_size_prop[] = {0, cpu_to_be32(spapr->htab_shift)};
> > > > > > +    sPAPRMachineClass *smc = 
> > > > > > SPAPR_MACHINE_GET_CLASS(qdev_get_machine());
> > > > > > +    sPAPRDRConnector *drc;
> > > > > > +    sPAPRDRConnectorClass *drck;
> > > > > > +    int drc_index;
> > > > > > +
> > > > > > +    if (smc->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)));
> > > > > > +    }
> > > > > >  
> > > > > >      /* Note: we keep CI large pages off for now because a 64K 
> > > > > > capable guest
> > > > > >       * provisioned with large pages might otherwise try to map a 
> > > > > > qemu
> > > > > > @@ -987,6 +999,16 @@ static void 
> > > > > > spapr_finalize_fdt(sPAPRMachineState *spapr,
> > > > > >          _FDT(spapr_drc_populate_dt(fdt, 0, NULL, 
> > > > > > SPAPR_DR_CONNECTOR_TYPE_LMB));
> > > > > >      }
> > > > > >  
> > > > > > +    if (smc->dr_cpu_enabled) {
> > > > > > +        int offset = fdt_path_offset(fdt, "/cpus");
> > > > > > +        ret = spapr_drc_populate_dt(fdt, offset, NULL,
> > > > > > +                                    SPAPR_DR_CONNECTOR_TYPE_CPU);
> > > > > > +        if (ret < 0) {
> > > > > > +            error_report("Couldn't set up CPU DR device tree 
> > > > > > properties");
> > > > > > +            exit(1);
> > > > > > +        }
> > > > > > +    }
> > > > > > +
> > > > > >      _FDT((fdt_pack(fdt)));
> > > > > >  
> > > > > >      if (fdt_totalsize(fdt) > FDT_MAX_SIZE) {
> > > > > > @@ -1181,7 +1203,7 @@ static void ppc_spapr_reset(void)
> > > > > >  
> > > > > >  }
> > > > > >  
> > > > > > -static void spapr_cpu_reset(void *opaque)
> > > > > > +void spapr_cpu_reset(void *opaque)
> > > > > >  {
> > > > > >      sPAPRMachineState *spapr = SPAPR_MACHINE(qdev_get_machine());
> > > > > >      PowerPCCPU *cpu = opaque;
> > > > > > @@ -1622,6 +1644,8 @@ static void spapr_boot_set(void *opaque, 
> > > > > > const char *boot_device,
> > > > > >  void spapr_cpu_init(sPAPRMachineState *spapr, PowerPCCPU *cpu, 
> > > > > > Error **errp)
> > > > > >  {
> > > > > >      CPUPPCState *env = &cpu->env;
> > > > > > +    CPUState *cs = CPU(cpu);
> > > > > > +    int i;
> > > > > >  
> > > > > >      /* Set time-base frequency to 512 MHz */
> > > > > >      cpu_ppc_tb_init(env, TIMEBASE_FREQ);
> > > > > > @@ -1646,6 +1670,14 @@ void spapr_cpu_init(sPAPRMachineState 
> > > > > > *spapr, PowerPCCPU *cpu, Error **errp)
> > > > > >          }
> > > > > >      }
> > > > > >  
> > > > > > +    /* 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;
> > > > > > +        }
> > > > > > +    }
> > > > > > +    
> > > > > 
> > > > > This hunk seems like it belongs in a different patch.    
> > > > 
> > > > It appears that this would be needed by other archs also to set the
> > > > NUMA node for the hot-plugged CPU. How about make an API out of this
> > > > and use this something like below ? Igor ?    
> > > 
> > > Is there a way we could put this in the the CPU thread initialization
> > > itself?  Rather than requiring every platform to call a helper.  
> > I'd suggest hotplugable CPU entity to have 'node' property, like we have
> > in pc-dimm.  
> 
> Ok.  Do you think that makes sense for the base core type (which will
> sometimes be a hotpluggable CPU entity and sometimes might not be)?
I wouldn't push it into generic base type, as in general target
implementing it in subtype might not have a numa concept at all.
I'd be better to leave decision up to target and corresponding subtype.

> 
> > However machine owns numa mapping, so setting it from thread
> > initialization seems to be wrong.  
> 
> Ah.. good point.
> 
> > Could that be done from machine's plug() handler (spapr_core_plug)?  
> 
> Yes, I think so.  The core might need a "set node" method, but we
> should be able to control this from the plug() handler.
> 
> > Also I notice that there is in no way to check/set 'address' properties
> > at machine level before calling cpu->realize(), which makes us
> > to attempt checking them inside cpu->realize() or at post realize time
> > in HotplugHandler(machine)->plug() hook with following full rollback
> > in case of failure.
> > Maybe we need add an HotplugHandler->pre_realize() hook to allow machine
> > verify/set properties of CPU before it's realized.  
> 
> I don't really see what the problem is with checking inside realize().
pls see http://patchwork.ozlabs.org/patch/594196/ for reasoning behind this