Re: [edk2-devel] [PATCH v2] OvmfPkg/PlatformInitLib: catch QEMU's CPU hotplug reg block regression

["Laszlo Ersek" <lersek@redhat.com>]

On 1/18/23 08:25, Gerd Hoffmann wrote:
> On Tue, Jan 17, 2023 at 05:43:53PM +0100, Ard Biesheuvel wrote:
>> On Tue, 17 Jan 2023 at 13:37, Gerd Hoffmann <kraxel@redhat.com> wrote:
>>>
>>>   Hi,
>>>
>>>>>> In particular the firmware makes no further decisions based on
>>>>>> whether QEMU advertized some of these features.
>>>>>
>>>>> I was thinking the other way around:  When cpu hotplug is disabled in
>>>>> qemu it should be safe to skip the whole cpu hotplug checking dance.
>>>>> See test patch below.
>>>>>
>>>>> That would give us a config switch (turn off cpu hotplug support)
>>>>> which would allow edk2 run on qemu versions with broken cpu hotplug.
>>>>>
>>>>> Does the idea look sane or do I miss something?
>>>
>>>> This would be wrong.
>>>>
>>>> [ detailed description snipped here (but stored for later reference,
>>>>   thanks for all the details) ]
>>>
>>> So, the tl;dr version:  cpu hotplug is older than smi feature
>>> negotiation, so smi hotplug feature bit being off doesn't imply
>>> qemu wouldn't hotplug cpus.
>>>
>>> So, no easy way out.  Luckily this affects tcg only.
>>>
>>> For edk2 ci doing (tcg) efi shell test boots switching to Oliver's
>>> latest containers with fixed qemu included should handle things
>>> (latest series just posted).  So once this is in we should be able to
>>> merge this patch without breaking CI.
>>
>> My head is spinning.
>>
>> What about running QEMU with only a single CPU, and without any of
>> these features? Is there really no way we can make that work without
>> turning OVMF into the timebomb that Laszlo describes?
> 
> I can't see any way :(
> 
> ovmf seeing only a single cpu does not imply cpu hotplug can't happen,
> it could be "qemu -smp cpus=1,maxcpus=4".  Figuring the maxcpus number
> depends on the broken cpu hotplug registers.
> 
>> It's just very annoying that on a non-KVM host and a given QEMU
>> binary, you might simply be out of luck entirely, and there is no way
>> you can run OVMF with the fix applied. I would like to avoid that if
>> possible.
> 
> Indeed.

... you could introduce a new fw_cfg boolean switch (and explain it in
the hang message) that meant: "I know what this QEMU bug is, I
understand its consequences are obscure, risky, and far-reaching in
OVMF, I've been warned, I know what I'm doing". That's a relatively
small addition to this patch, and then the risk is assumed by the user.
It resolves "being out of luck *entirely*".

Laszlo