Re: [RFC 3/3] x68: acpi: trigger SMI before scanning for hotplugged CPUs

[Laszlo Ersek <lersek@redhat.com>]

On 07/14/20 17:19, Igor Mammedov wrote:
> On Tue, 14 Jul 2020 14:41:28 +0200
> Laszlo Ersek <lersek@redhat.com> wrote:
> 
>> On 07/14/20 14:28, Laszlo Ersek wrote:
>>> (CC'ing Peter Krempa due to virsh setvcpu (singular) / setvcpus (plural)
>>> references)
>>>
>>> On 07/10/20 18:17, Igor Mammedov wrote:  
>>>> In case firmware has negotiated CPU hotplug SMI feature, generate
>>>> AML to describe SMI IO port region and send SMI to firmware
>>>> on each CPU hotplug SCI.
>>>>
>>>> It might be not really usable, but should serve as a starting point to
>>>> discuss how better to deal with split hotplug sequence during hot-add
>>>> (
>>>> ex scenario where it will break is:
>>>>    hot-add  
>>>>       -> (QEMU) add CPU in hotplug regs
>>>>       -> (QEMU) SCI  
>>>>            -1-> (OS) scan
>>>>                -1-> (OS) SMI
>>>>                -1-> (FW) pull in CPU1 ***
>>>>                -1-> (OS) start iterating hotplug regs
>>>>    hot-add  
>>>>       -> (QEMU) add CPU in hotplug regs
>>>>       -> (QEMU) SCI  
>>>>             -2-> (OS) scan (blocked on mutex till previous scan is finished)
>>>>                -1-> (OS) 1st added CPU1 send device check event -> INIT/SIPI
>>>>                -1-> (OS) 1st added CPU2 send device check event -> INIT/SIPI
>>>>                        that's where it explodes, since FW didn't see CPU2
>>>>                        when SMI was called
>>>> )
>>>>
>>>> hot remove will throw in yet another set of problems, so lets discuss
>>>> both here and see if we can  really share hotplug registers block between
>>>> FW and AML or we should do something else with it.  
>>>
>>> This issue is generally triggered by management applications such as
>>> libvirt that issue device_add commands in quick succession. For libvirt,
>>> the command is "virsh setvcpus" (plural) with multiple CPUs specified
>>> for plugging. The singular "virsh setvcpu" command, which is more
>>> friendly towards guest NUMA, does not run into the symptom.
>>>
>>> The scope of the scan method lock is not large enough, with SMI in the
>>> picture.
>>>
>>> I suggest that we not uproot the existing AML code or the hotplug
>>> register block. Instead, I suggest that we add serialization at a higher
>>> level, with sufficient scope.
>>>
>>> QEMU:
>>>
>>> - introduce a new flag standing for "CPU plug operation in progress"
>>>
>>> - if ICH9_LPC_SMI_F_BROADCAST_BIT has been negotiated:
>>>
>>>   - "device_add" and "device_del" should enforce
>>>     ICH9_LPC_SMI_F_CPU_HOTPLUG_BIT and
>>>     ICH9_LPC_SMI_F_CPU_HOT_UNPLUG_BIT, respectively
>>>
>>>   - both device_add and device_del (for VCPUs) should set check the
>>>     "in progress" flag.
>>>
>>>     - If set, reject the request synchronously
>>>
>>>     - Otherwise, set the flag, and commence the operation
>>>
>>>   - in cpu_hotplug_wr(), where we emit the ACPI_DEVICE_OST event with
>>>     qapi_event_send_acpi_device_ost(), clear the "in-progress" flag.
>>>
>>> - If QEMU executes the QMP command processing and the cpu_hotplug_wr()
>>> function on different (host OS) threads, then perhaps we should use an
>>> atomic type for the flag. (Not sure about locking between QEMU threads,
>>> sorry.) I don't really expect race conditions, but in case we ever get
>>> stuck with the flag, we should make sure that the stuck state is "in
>>> progress", and not "not in progress". (The former state can prevent
>>> further plug operations, but cannot cause the guest to lose state.)  
>>
>> Furthermore, the "CPU plug operation in progress" flag should be:
>> - either migrated,
>> - or a migration blocker.
>>
>> Because on the destination host, device_add should be possible if and
>> only if the plug operation completed (either still on the source host,
>> or on the destination host).
> 
> I have a way more simple alternative idea, which doesn't involve libvirt.
> 
> We can change AML to
>   1. cache hotplugged CPUs from controller
>   2. send SMI
>   3. send Notify event to OS to online cached CPUs
> this way we never INIT/SIPI cpus that FW hasn't seen yet
> as for FW, it can relocate extra CPU that arrived after #1
> it won't cause any harm as on the next SCI AML will pick up those
> CPUs and SMI upcall will be just NOP.
> 
> I'll post a patch here on top of this series for you to try
> (without any of your comments addressed yet, as it's already written
> and I was testing it for a while to make sure it won't explode
> with various windows versions)

Sounds good, I'll be happy to test it.

Indeed "no event" is something that the fw deals with gracefully. (IIRC
I wanted to cover a "spurious SMI" explicitly.)

It didn't occur to me that you could dynamically size e.g. a package
object in AML. Based on my reading of the ACPI spec, "VarPackageOp" can
take a *runtime* "NumElements", so if you did two loops, the first loop
could count the pending stuff, and then a VarPackageOp could be used
with just the right NumElements... Anyway, I digress :)

> 
>> I guess that the "migration blocker" option is easier.
>>
>> Anyway I assume this is already handled with memory hotplug somehow
>> (i.e., migration attempt between device_add and ACPI_DEVICE_OST).
> 
> Thanks for comments,
> I'll need some time to ponder on other comments and do some
> palaeontology research to answer questions
> (aka. I need to make up excuses for the code I wrote :) )

haha, thanks :)
Laszlo