Re: [PATCH for-6.0 4/8] spapr/xive: Add "nr-ipis" property

[Cédric Le Goater]

On 11/20/20 6:46 PM, Greg Kurz wrote:
> The sPAPR XIVE device exposes a range of LISNs that the guest uses
> for IPIs. This range is currently sized according to the highest
> vCPU id, ie. spapr_max_server_number(), as obtained from the machine
> through the "nr_servers" argument of the generic spapr_irq_dt() call.
> 
> This makes sense for the "ibm,interrupt-server-ranges" property of
> sPAPR XICS, but certainly not for "ibm,xive-lisn-ranges". The range
> should be sized to the maximum number of possible vCPUs. It happens
> that spapr_max_server_number() == smp.max_cpus in practice with the
> machine default settings. This might not be the case though when
> VSMT is in use : we can end up with a much larger range (up to 8
> times bigger) than needed and waste LISNs. 

will it exceed 4K ? if not, we are fine.

The fact that it is so complex to get the number of vCPUs is a 
problem by it self to me. Can we simplify that ? or introduce a 
spapr_max_server_number_id() ? 

> But most importantly, this
> lures people into thinking that IPI numbers are always equal to
> vCPU ids, which is wrong and bit us recently:

do we have a converting routing vcpu_id_to_ipi ? I think we have
that in KVM.

> https://lists.gnu.org/archive/html/qemu-devel/2020-11/msg01476.html
> 
> Introduce an "nr-ipis" property that the machine sets to smp.max_cpus
> before realizing the deice. Use that instead of "nr_servers" in
> spapr_xive_dt(). Have the machine to claim the same number of IPIs
> in spapr_irq_init().
> 
> This doesn't cause any guest visible change when using the machine
> default settings (ie. VSMT == smp.threads).>
> Signed-off-by: Greg Kurz <groug@kaod.org>
> ---
>  include/hw/ppc/spapr_xive.h | 8 ++++++++
>  hw/intc/spapr_xive.c        | 4 +++-
>  hw/ppc/spapr_irq.c          | 4 +++-
>  3 files changed, 14 insertions(+), 2 deletions(-)
> 
> diff --git a/include/hw/ppc/spapr_xive.h b/include/hw/ppc/spapr_xive.h
> index 7123ea07ed78..69b9fbc1b9a5 100644
> --- a/include/hw/ppc/spapr_xive.h
> +++ b/include/hw/ppc/spapr_xive.h
> @@ -43,6 +43,14 @@ typedef struct SpaprXive {
>  
>      /* DT */
>      gchar *nodename;
> +    /*
> +     * The sPAPR XIVE device needs to know how many vCPUs it
> +     * might be exposed to in order to size the IPI range in
> +     * "ibm,interrupt-server-ranges". It is the purpose of the

There is no "ibm,interrupt-server-ranges"  property in XIVE

> +     * "nr-ipis" property which *must* be set to a non-null
> +     * value before realizing the sPAPR XIVE device.
> +     */
> +    uint32_t nr_ipis;
>  
>      /* Routing table */
>      XiveEAS       *eat;
> diff --git a/hw/intc/spapr_xive.c b/hw/intc/spapr_xive.c
> index e4dbf9c2c409..d13a2955ce9b 100644
> --- a/hw/intc/spapr_xive.c
> +++ b/hw/intc/spapr_xive.c
> @@ -311,6 +311,7 @@ static void spapr_xive_realize(DeviceState *dev, Error 
> **errp)
>      /* Set by spapr_irq_init() */
>      g_assert(xive->nr_irqs);
>      g_assert(xive->nr_servers);
> +    g_assert(xive->nr_ipis);
>  
>      sxc->parent_realize(dev, &local_err);
>      if (local_err) {
> @@ -608,6 +609,7 @@ static Property spapr_xive_properties[] = {
>       */
>      DEFINE_PROP_UINT32("nr-ends", SpaprXive, nr_ends_vmstate, 0),
>      DEFINE_PROP_UINT32("nr-servers", SpaprXive, nr_servers, 0),
> +    DEFINE_PROP_UINT32("nr-ipis", SpaprXive, nr_ipis, 0),
>      DEFINE_PROP_UINT64("vc-base", SpaprXive, vc_base, SPAPR_XIVE_VC_BASE),
>      DEFINE_PROP_UINT64("tm-base", SpaprXive, tm_base, SPAPR_XIVE_TM_BASE),
>      DEFINE_PROP_UINT8("hv-prio", SpaprXive, hv_prio, 7),
> @@ -698,7 +700,7 @@ static void spapr_xive_dt(SpaprInterruptController *intc, 
> uint32_t nr_servers,
>      /* Interrupt number ranges for the IPIs */
>      uint32_t lisn_ranges[] = {
>          cpu_to_be32(SPAPR_IRQ_IPI),
> -        cpu_to_be32(SPAPR_IRQ_IPI + nr_servers),
> +        cpu_to_be32(SPAPR_IRQ_IPI + xive->nr_ipis),

I don't understand why we need nr_ipis. Can't we pass the same value in 
nr_servers from the machine ?

( Conceptually, the first 4K are all IPIs. The first range is for 
  HW threads ) 


>      };
>      /*
>       * EQ size - the sizes of pages supported by the system 4K, 64K,
> diff --git a/hw/ppc/spapr_irq.c b/hw/ppc/spapr_irq.c
> index 8c5627225636..a0fc474ecb06 100644
> --- a/hw/ppc/spapr_irq.c
> +++ b/hw/ppc/spapr_irq.c
> @@ -325,12 +325,14 @@ void spapr_irq_init(SpaprMachineState *spapr, Error 
> **errp)
>  
>      if (spapr->irq->xive) {
>          uint32_t nr_servers = spapr_max_server_number(spapr);
> +        uint32_t max_cpus = MACHINE(spapr)->smp.max_cpus;

So that's the value we should be using in case of VSMT and not 
spapr_max_server_number(). If I am correct, this is a max_cpu_id ?


>          DeviceState *dev;
>          int i;
>  
>          dev = qdev_new(TYPE_SPAPR_XIVE);
>          qdev_prop_set_uint32(dev, "nr-irqs", smc->nr_xirqs + 
> SPAPR_XIRQ_BASE);
>          qdev_prop_set_uint32(dev, "nr-servers", nr_servers);
> +        qdev_prop_set_uint32(dev, "nr-ipis", max_cpus);
>          object_property_set_link(OBJECT(dev), "xive-fabric", OBJECT(spapr),
>                                   &error_abort);
>          sysbus_realize_and_unref(SYS_BUS_DEVICE(dev), &error_fatal);
> @@ -338,7 +340,7 @@ void spapr_irq_init(SpaprMachineState *spapr, Error 
> **errp)
>          spapr->xive = SPAPR_XIVE(dev);
>  
>          /* Enable the CPU IPIs */
> -        for (i = 0; i < nr_servers; ++i) {
> +        for (i = 0; i < max_cpus; ++i) {
>              SpaprInterruptControllerClass *sicc
>                  = SPAPR_INTC_GET_CLASS(spapr->xive);