[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [Qemu-ppc] [PATCH] spapr: Add H-Call H_HOME_NODE_ASSOCI
From: |
Greg Kurz |
Subject: |
Re: [Qemu-devel] [Qemu-ppc] [PATCH] spapr: Add H-Call H_HOME_NODE_ASSOCIATIVITY |
Date: |
Tue, 18 Dec 2018 11:50:15 +0100 |
On Tue, 18 Dec 2018 11:00:01 +0100
Laurent Vivier <address@hidden> wrote:
> On 18/12/2018 10:23, Greg Kurz wrote:
> > On Tue, 18 Dec 2018 08:50:00 +0100
> > Laurent Vivier <address@hidden> wrote:
> >
> >> On 18/12/2018 05:29, David Gibson wrote:
> >>> On Mon, Dec 17, 2018 at 03:00:55PM +0100, Laurent Vivier wrote:
> >>>> H_HOME_NODE_ASSOCIATIVITY H-Call returns the associativity domain
> >>>> designation associated with the identifier input parameter.
> >>>>
> >>>> Remove the warning message from the kernel:
> >>>> VPHN is not supported. Disabling polling..
> >>>>
> >>>> Signed-off-by: Laurent Vivier <address@hidden>
> >>>
> >>> From the looks of PAPR, I suspect this call isn't of much use outside
> >>> PowerVM guests, though it probably wouldn't do any harm.
> >>
> >> This call is used by the kernel to get the node id of a CPU on hotplug
> >> and fixes a crash when we hotplug a CPU in a memory-less/CPU-less node
> >> where this information is missing (not initialized from the device-tree).
> >>
> >
> > So this patch isn't just about removing the warning message from the kernel
> > but about fixing an actual crash ?
>
> Yes, I updated the message but sent the wrong e-mail.
>
> > I ask because if it's only about the warning, why does the kernel call
> > H_HOME_NODE_ASSOCIATIVITY when hcall-vphn isn't advertised ? Especially,
> > the polling for topology changes is only started if hcall-vphn is present:
> >
> > if (firmware_has_feature(FW_FEATURE_VPHN) &&
> > lppaca_shared_proc(get_lppaca())) {
> > if (!vphn_enabled) {
> > vphn_enabled = 1;
> > setup_cpu_associativity_change_counters();
> > timer_setup(&topology_timer, topology_timer_fn,
> > TIMER_DEFERRABLE);
> > reset_topology_timer();
> > }
> > }
> >
> > It thus seems wrong to emit the "Disable polling.." warning for something
> > that was never enabled in the first place, doesn't it ?
>
> It's unconditionally called from find_and_online_cpu_nid() that is used
> to plug a CPU in a node that is not already online.
>
> > On the other hand, if this really needed to avoid a crash, I guess you
> > should provide some more details.
>
> I agree.
>
> >>> BenH, Paulus, any thoughts?
> >>>
> >>> One nit in implementation: if you implement this hcall, it's supposed
> >>> to be advertised by adding hcall-vphn to ibm,hypertas-functions.
> >> ok in v2.
> >>
> >> Thanks,
> >> Laurent
> >>
> >>>> ---
> >>>> Based-on: <address@hidden>
> >>>> "[PULL 00/27] ppc-for-4.0 queue 20181213"
> >>>>
> >>>> hw/ppc/spapr_hcall.c | 39 +++++++++++++++++++++++++++++++++++++++
> >>>> include/hw/ppc/spapr.h | 1 +
> >>>> 2 files changed, 40 insertions(+)
> >>>>
> >>>> diff --git a/hw/ppc/spapr_hcall.c b/hw/ppc/spapr_hcall.c
> >>>> index 78fecc8fe9..454ec594fd 100644
> >>>> --- a/hw/ppc/spapr_hcall.c
> >>>> +++ b/hw/ppc/spapr_hcall.c
> >>>> @@ -1663,6 +1663,41 @@ static target_ulong
> >>>> h_client_architecture_support(PowerPCCPU *cpu,
> >>>> return H_SUCCESS;
> >>>> }
> >>>>
> >>>> +static target_ulong h_home_node_associativity(PowerPCCPU *cpu,
> >>>> + sPAPRMachineState *spapr,
> >>>> + target_ulong opcode,
> >>>> + target_ulong *args)
> >>>> +{
> >>>> + target_ulong flags = args[0];
> >>>> + target_ulong procno = args[1];
> >>>> + PowerPCCPU *tcpu;
> >>>> + int idx;
> >>>> +
> >>>> + /* only support procno from H_REGISTER_VPA */
> >>>> + if ((flags & 0x1) == 0) {
> >>>> + return H_PARAMETER;
> >>>> + }
> >
> > LoPAPR says that the guest can pass exactly 0x1 or 0x2 in flags. The
> > above check should then rather be flags == 0x1.
> >
>
> ok
>
> > Also, even if linux only seems to call this with 0x1, this is a
> > limitation from a LoPAPR standpoint. Not sure H_PARAMETER is the
> > appropriate return value if flags is 0x2 since the guest did
> > nothing wrong... I'd rather return H_FUNCTION in this case.
>
> The doc says:
>
> H_Function: The function is not supported
> H_Parameter: Unsupported flag parameter value
>
> in that case function is supported but not the flag, so I think
> H_PARAMETER is a better choice.
>
Well... neither LoPAPR, nor IBM confidential PAPR+ do say anything
about partial support for this hcall. If the guest was to use the
flags == 0x2 variant, eg, some closed-source OS supporting PAPR,
it could be legitimately confused to get an H_PARAMETER error when
passing supposedly valid parameters... how to cope with that ?
On the other hand, if QEMU cannot honor the flags == 0x2 variant
and returns H_FUNCTION then the OS can recover since it is
required by the specification.
But I don't really care for now, and we can talk about this later
if I'm assigned a BZ from the people who run such OS in KVM guests ;)
> Thanks,
> Laurent
>