Re: [PATCH for-6.0 2/8] spapr/xive: Introduce spapr_xive_nr_ends()

[Cédric Le Goater]

On 11/24/20 6:01 PM, Greg Kurz wrote:
> On Tue, 24 Nov 2020 14:54:38 +0100
> Cédric Le Goater <clg@kaod.org> wrote:
> 
>> On 11/23/20 12:16 PM, Greg Kurz wrote:
>>> On Mon, 23 Nov 2020 10:46:38 +0100
>>> Cédric Le Goater <clg@kaod.org> wrote:
>>>
>>>> On 11/20/20 6:46 PM, Greg Kurz wrote:
>>>>> We're going to kill the "nr_ends" field in a subsequent patch.
>>>>
>>>> why ? it is one of the tables of the controller and its part of 
>>>> the main XIVE concepts. Conceptually, we could let the machine 
>>>> dimension it with an arbitrary value as OPAL does. The controller
>>>> would fail when the table is fully used. 
>>>>
>>>
>>> The idea is that the sPAPR machine only true need is to create a
>>> controller that can accommodate up to a certain number of vCPU ids.
>>> It doesn't really to know about the END itself IMHO.> 
>>> This being said, if we decide to pass both spapr_max_server_number()
>>> and smp.max_cpus down to the backends as function arguments, we won't
>>> have to change "nr_ends" at all.
>>
>> I would prefer that but I am still not sure what they represent. 
>>
>> Looking at the sPAPR XIVE code, we deal with numbers/ranges in the 
>> following places today.
>>
>>  * spapr_xive_dt() 
>>
>>    It defines a range of interrupt numbers to be used by the guest 
>>    for the threads/vCPUs IPIs. It's a subset of interrupt numbers 
>>    in :
>>
>>              [ SPAPR_IRQ_IPI - SPAPR_IRQ_EPOW [
>>
>>    These are not vCPU ids.
>>
>>    Since these interrupt numbers will be considered as free to use
>>    by the OS, it makes sense to pre-claim them. But claiming an 
>>    interrupt number in the guest can potentially set up, through 
>>    the KVM device, a mapping on the host and in HW. See below why
>>    this can be a problem.
>>
>>  * kvmppc_xive_cpu_connect()   
>>
>>    This sizes the NVT tables in OPAL for the guest. This is the  
>>    max number of vCPUs of the guest (not vCPU ids)
>>
> 
> I guess you're talking about KVM_DEV_XIVE_NR_SERVERS in
> kvmppc_xive_connect() actually. We're currently passing
> spapr_max_server_number() (vCPU id) but you might be
> right.
> 
> I need to re-read the story around VSMT and XIVE.

ok. What we care about here, is a size to allocate the NVT block
representing the vCPUs in HW. NVT ids are pushed in the thread 
contexts when the vCPUs are scheduled to run and looked for by 
the presenter when an interrupt is to be delivered.

 
> commit 1e175d2e07c71d9574f5b1c74523abca54e2654f
> Author: Sam Bobroff <sam.bobroff@au1.ibm.com>
> Date:   Wed Jul 25 16:12:02 2018 +1000
> 
>     KVM: PPC: Book3S HV: Pack VCORE IDs to access full VCPU ID space
> 
>>  * spapr_irq_init()
>>
>>    This is where the IPI interrupt numbers are claimed today. 
>>    Directly in QEMU and KVM, if the machine is running XIVE only, 
>>    indirectly if it's dual, first in QEMU and then in KVM when 
>>    the machine switches of interrupt mode in CAS. 
>>
>>    The problem is that the underlying XIVE resources in HW are 
>>    allocated where the QEMU process is running. Which is not the
>>    best option when the vCPUs are pinned on different chips.
>>
>>    My patchset was trying to improve that by claiming the IPI on 
>>    demand when the vCPU is connected to the KVM device. But it 
>>    was using the vCPU id as the IPI interrupt number which is 
>>    utterly wrong, the guest OS could use any number in the range 
>>    exposed in the DT.
>>    
>>    The last patch you sent was going in the right direction I think.
>>    That is to claim the IPI when the guest OS is requesting for it. 
>>
>>    
>> 160528045027.804522.6161091782230763832.stgit@bahia.lan/">http://patchwork.ozlabs.org/project/qemu-devel/patch/160528045027.804522.6161091782230763832.stgit@bahia.lan/
>>    
>>    But I don't understand why it was so complex. It should be like
>>    the MSIs claimed by PCI devices.
>>
> 
> The difference here is that the guest doesn't claim IPIs. They are
> supposedly pre-claimed in "ibm,xive-lisn-ranges". And this is actually
> the case in QEMU.

yes. That's what I want to change (for performance)

> The IPI setup sequence in the guest is basically:
> 1) grab a free irq from the bitmap, ie. "ibm,xive-lisn-ranges"
> 2) calls H_INT_GET_SOURCE_INFO, ie. populate_irq_data()
> 3) calls H_INT_SET_SOURCE_CONFIG, ie, configure_irq())
> 
> If we want an IPI to be claimed by the appropriate vCPU, we
> can only do this from under H_INT_SET_SOURCE_CONFIG. And
> until the guest eventually configures the IPI, KVM and QEMU
> are out of sync.

Well, KVM doesn't know either how much PCI MSIs will be claimed.
It all depends on the guest OS. 

I don't think this is a problem to expose unclaimed interrupt
numbers to the guest if they are IPIs. We can detect that
easily with the range and claim the source at KVM level when 
it's configured or in h_int_get_source_info(). Talking of which, 
it might be good to have a KVM command to query the source 
characteristics on the host. I sent a patchset a while ago in 
that sense.

> This complexifies migration because we have to guess at
> post load if we should claim the IPI in KVM or not. The
> simple presence of the vCPU isn't enough : we need to
> guess if the guest actually configured the IPI or not.

The EAT will be transferred from the source and the call to 
kvmppc_xive_source_reset_one() should initialize the KVM 
device correctly on the target for all interrupts.

>> All this to say, that we need to size better the range in the 
>> "ibm,xive-lisn-ranges" property if that's broken for vSMT. 
>>
> 
> Sizing the range to smp.max_cpus as proposed in this series
> is fine, no matter what the VSMT is.

ok. That's a fix for spapr_irq_dt() then. And possibly, there 
is a similar one for KVM_DEV_XIVE_NR_SERVERS.
 
>> Then, I think the IPIs can be treated just like the PCI MSIs
>> but they need to be claimed first. That's the ugly part. 
>>
> 
> Yeah that's the big difference. For PCI MSIs, QEMU owns the
> bitmap and the guest can claim (or release) a number of
> MSIs the "ibm,change-msi" RTAS interface. There's no
> such thing for IPIs : they are supposedly already claimed.

IPIs are a bit special because there are no I/O devices to
claim them. We could consider the vCPU has being the device. 
That was my first attempt but it was wrong since the OS is 
in charge of choosing an interrupt number for the IPI. 

>> Should we add a special check in h_int_set_source_config to
>> deal with unclaimed IPIs that are being configured ?
>>
> 
> This is what my tentative fix does.

I didn't understand the complexity of it, may be due to my
patchset.

You should try again :)

C.