Re: [Qemu-devel] [PATCH] hw/pci-host: Fix x86 Host Bridges 64bit PCI hole

[Marcel Apfelbaum <marcel@redhat.com>]

On 20/10/2017 13:59, Gerd Hoffmann wrote:
> On Fri, 2017-10-20 at 11:32 +0200, Laszlo Ersek wrote:
>> On 10/20/17 08:55, Gerd Hoffmann wrote:
>>>    Hi,
>>>
>>>>> commit message says:
>>>>>
>>>>> <quote>
>>>>>       It turns out that some 32 bit windows guests crash
>>>>>       if 64 bit PCI hole size is >2G.
>>>>> </quote>
>>>>>
>>>>> Why this suddenly isn't a problem any more?
>>>>>
>>>>
>>>> I suppose it is, so we need a way to turn it "off".
>>>
>>> Or have machine types behave differently, i.e. give q35 a large
>>> 64bit
>>> hole and leave pc as-is.
>>
>> *If* we make it dependent on machine types at all, then please also
>> make
>> it versioned for Q35.
> 
> We probably need that for live migration compatibility anyway.  If we
> add a switch we can add it to both pc and q35 (I think they share most
> code so little extra cost), but have different defaults depending on
> machine type.
> 
>>> BTW:  Is it safe to just assume 40 bits physical is going to
>>> work?  My
>>> workstation:
>>>
>>> model name      : Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz
>>> address sizes   : 39 bits physical, 48 bits virtual
>>>
>>> Does this imply ept is limited 39 bits physical too?
>>
>> Very good point to raise; "39 bits physical" on your workstation
>> *does*
>> imply that EPT is limited exactly the same way. I had run into this
>> very
>> problem while working on 64GB+ RAM enablement in OVMF. (Back then my
>> laptop had: "address sizes : 36 bits physical, 48 bits virtual".)
> 
> Ah, right, I remember.  This is why we ended up with 32G window 32G
> aligned above highest memory (or reserved) address.  No address above
> 64G will be used unless you have lots of memory in your virtual machine
> (in which case your host very likely supports more than 36 address
> lines).
> 
> So I guess the options are to play safe and do something simliar on the
> qemu side, or go figure how much physical address space is available
> and use that (which should also solve the "what do we with >1TB guests"
> issue).  The later could cause some interesting live migration issues
> though, when migrating between hosts with different physical address
> space sizes.
> 
> So maybe the pci-hole64-size option isn't that bad after all?  We could
> make it default to 2G on pc and 32G on q35 ...
> 

I could live with that, as I previously stated I started that way.
The property is already there, if anyone needs a bigger 64bit PCI
hole, libvirt can set a new value.
Adding Laine to give him the opportunity to object :)

I'll come up with V3 that actually implements the pci-hole64-size,
having *some* 64bit window for PCI hotplug is better than none.

Thanks all for the help!
Marcel

> cheers,
>    Gerd
>