Re: [RFC 0/3] acpi: cphp: add CPHP_GET_CPU_ID_CMD command to cpu hotplug MMIO interface

[Laszlo Ersek <lersek@redhat.com>]

On 10/10/19 21:20, Eduardo Habkost wrote:
> On Thu, Oct 10, 2019 at 05:57:54PM +0200, Igor Mammedov wrote:
>> On Thu, 10 Oct 2019 09:59:42 -0400
>> "Michael S. Tsirkin" <mst@redhat.com> wrote:
>>
>>> On Thu, Oct 10, 2019 at 03:39:12PM +0200, Igor Mammedov wrote:
>>>> On Thu, 10 Oct 2019 05:56:55 -0400
>>>> "Michael S. Tsirkin" <mst@redhat.com> wrote:
>>>>
>>>>> On Wed, Oct 09, 2019 at 09:22:49AM -0400, Igor Mammedov wrote:
>>>>>> As an alternative to passing to firmware topology info via new fwcfg files
>>>>>> so it could recreate APIC IDs based on it and order CPUs are enumerated,
>>>>>>
>>>>>> extend CPU hotplug interface to return APIC ID as response to the new command
>>>>>> CPHP_GET_CPU_ID_CMD.  
>>>>>
>>>>> One big piece missing here is motivation:
>>>> I thought the only willing reader was Laszlo (who is aware of context)
>>>> so I skipped on details and confused others :/
>>>>
>>>>> Who's going to use this interface?
>>>> In current state it's for firmware, since ACPI tables can cheat
>>>> by having APIC IDs statically built in.
>>>>
>>>> If we were creating CPU objects in ACPI dynamically
>>>> we would be using this command as well.
>>>
>>> I'm not sure how it's even possible to create devices dynamically. Well
>>> I guess it's possible with LoadTable. Is this what you had in
>>> mind?
>>
>> Yep. I even played this shiny toy and I can say it's very tempting one.
>> On the  other side, even problem of legacy OSes not working with it aside,
>> it's hard to debug and reproduce compared to static tables.
>> So from maintaining pov I dislike it enough to be against it.
>>
>>
>>>> It would save
>>>> us quite a bit space in ACPI blob but it would be a pain
>>>> to debug and diagnose problems in ACPI tables, so I'd rather
>>>> stay with static CPU descriptions in ACPI tables for the sake
>>>> of maintenance.
>>>>> So far CPU hotplug was used by the ACPI, so we didn't
>>>>> really commit to a fixed interface too strongly.
>>>>>
>>>>> Is this a replacement to Laszlo's fw cfg interface?
>>>>> If yes is the idea that OVMF going to depend on CPU hotplug directly then?
>>>>> It does not depend on it now, does it?
>>>> It doesn't, but then it doesn't support cpu hotplug,
>>>> OVMF(SMM) needs to cooperate with QEMU "and" ACPI tables to perform
>>>> the task and using the same interface/code path between all involved
>>>> parties makes the task easier with the least amount of duplicated
>>>> interfaces and more robust.
>>>>
>>>> Re-implementing alternative interface for firmware (fwcfg or what not)
>>>> would work as well, but it's only question of time when ACPI and
>>>> this new interface disagree on how world works and process falls
>>>> apart.
>>>
>>> Then we should consider switching acpi to use fw cfg.
>>> Or build another interface that can scale.
>>
>> Could be an option, it would be a pain to write a driver in AML for fwcfg access though
>> (I've looked at possibility to access fwcfg from AML about a year ago and gave up.
>> I'm definitely not volunteering for the second attempt and can't even give an estimate
>> it it's viable approach).
>>
>> But what scaling issue you are talking about, exactly?
>> With current CPU hotplug interface we can handle upto UNIT32_MAX cpus, and extend
>> interface without need to increase IO window we are using now.
>>
>> Granted IO access it not fastest compared to fwcfg in DMA mode, but we already
>> doing stop machine when switching to SMM which is orders of magnitude slower.
>> Consensus was to compromise on speed of CPU hotplug versus more complex and more
>> problematic unicast SMM mode in OVMF (can't find a particular email but we have discussed
>> it with Laszlo already, when I considered ways to optimize hotplug speed)
> 
> If we were designing the interface from the ground up, I would
> agree with Michael.  But I don't see why we would reimplement
> everything from scratch now, if just providing the
> cpu_selector => cpu_hardware_id mapping to firmware is enough to
> make the existing interface work.
> 
> If somebody is really unhappy with the current interface and
> wants to implement a new purely fw_cfg-based one (and write the
> corresponding ACPI code), they would be welcome.

Let me re-iterate the difficulties quickly:

- DMA-based fw_cfg is troublesome in SEV guests (do you want to mess
with page table entries in AML methods? or pre-allocate an always
decrypted opregion? how large?)

- IO port based fw_cfg does not support writes (and I reckon that, when
the *OS* handles a hotplug event, it does have to talk back to QEMU)

- the CPU hotplug AML would have to arbitrate with Linux's own fw_cfg
driver (which exposes fw_cfg files to userspace, yay! /s)

In the phys world, CPU hotplug takes dedicated RAS hardware. Shoehorning
CPU hotplug into *firmware* config, when in two use cases [*], the
firmware shouldn't even know about CPU hotplug, feels messy.

[*] being (a) SeaBIOS, and (b) OVMF built without SMM

> I just don't see why we should spend our time doing that now.

I have to agree, we're already spread thin.

... I must admit: I didn't expect this, but now I've grown to *prefer*
the CPU hotplug register block!

Laszlo