Virtualization - Mar 2020 - [PATCH v2 1/3] iommu/virtio: Add topology description to virtio-iommu config space

On Tue, Mar 03, 2020 at 09:00:05AM -0500, Michael S. Tsirkin
wrote:
> Not necessarily. E.g. some power systems have neither.
> There are also systems looking to bypass ACPI e.g. for boot speed.
If there is no firmware layer between the hardware and the OS the
necessary information the OS needs to run on the hardware is probably
hard-coded into the kernel? In that case the same can be done with
virtio-iommu tolopology.
> That sentence doesn't really answer the question, does it?
To be more elaborate, putting this information into config space is a
layering violation. Hardware is never completly self-descriptive and
that is why there is the firmware which provides the information about
the hardware to the OS in a generic way.
> Frankly with platform specific interfaces like ACPI, virtio-iommu is
> much less compelling.  Describing topology as part of the device in a
> way that is first, portable, and second, is a good fit for hypervisors,
> is to me one of the main reasons virtio-iommu makes sense at all.
Virtio-IOMMU makes sense in the first place because it is much faster
than emulating one of the hardware IOMMUs. And an ACPI table is also
portable to all ACPI platforms, same with device-tree.

Regards,

	Joerg

On Tue, Mar 03, 2020 at 04:53:19PM +0100, Joerg Roedel
wrote:
> On Tue, Mar 03, 2020 at 09:00:05AM -0500, Michael S. Tsirkin wrote:
> > Not necessarily. E.g. some power systems have neither.
> > There are also systems looking to bypass ACPI e.g. for boot speed.
> 
> If there is no firmware layer between the hardware and the OS the
> necessary information the OS needs to run on the hardware is probably
> hard-coded into the kernel?
No. It's coded into the hardware. Which might even be practical
for bare-metal (e.g. on-board flash), but is very practical
when the device is part of a hypervisor.
> In that case the same can be done with
> virtio-iommu tolopology.
> 
> > That sentence doesn't really answer the question, does it?
> 
> To be more elaborate, putting this information into config space is a
> layering violation. Hardware is never completly self-descriptive

This "hardware" is actually part of hypervisor so there's no
reason it can't be completely self-descriptive. It's specified
by the same entity as the "firmware".

> and
> that is why there is the firmware which provides the information about
> the hardware to the OS in a generic way.
>
> > Frankly with platform specific interfaces like ACPI, virtio-iommu is
> > much less compelling.  Describing topology as part of the device in a
> > way that is first, portable, and second, is a good fit for
hypervisors,
> > is to me one of the main reasons virtio-iommu makes sense at all.
> 
> Virtio-IOMMU makes sense in the first place because it is much faster
> than emulating one of the hardware IOMMUs.
I don't see why it would be much faster. The interface isn't that
different from command queues of VTD.
> And an ACPI table is also
> portable to all ACPI platforms, same with device-tree.
> 
> Regards,
> 
> 	Joerg
Making ACPI meet the goals of embedded projects such as kata containers
would be a gigantic task with huge stability implications.  By
comparison this 400-line parser is well contained and does the job.  I
didn't yet see compelling reasons not to merge this, but I'll be
interested to see some more specific concerns.


-- 
MST

Hi Michael, Joerg,

On 3/3/20 5:09 PM, Michael S. Tsirkin wrote:
> On Tue, Mar 03, 2020 at 04:53:19PM +0100, Joerg Roedel wrote:
>> On Tue, Mar 03, 2020 at 09:00:05AM -0500, Michael S. Tsirkin wrote:
>>> Not necessarily. E.g. some power systems have neither.
>>> There are also systems looking to bypass ACPI e.g. for boot speed.
>>
>> If there is no firmware layer between the hardware and the OS the
>> necessary information the OS needs to run on the hardware is probably
>> hard-coded into the kernel?
> 
> No. It's coded into the hardware. Which might even be practical
> for bare-metal (e.g. on-board flash), but is very practical
> when the device is part of a hypervisor.
> 
>> In that case the same can be done with
>> virtio-iommu tolopology.
>>
>>> That sentence doesn't really answer the question, does it?
>>
>> To be more elaborate, putting this information into config space is a
>> layering violation. Hardware is never completly self-descriptive
> 
> 
> This "hardware" is actually part of hypervisor so there's no
> reason it can't be completely self-descriptive. It's specified
> by the same entity as the "firmware".
Yes in the QEMU case the machine code would fill this information as it
knows all the devices connected to the iommu. In the same way it would
generate the DT description or the ACPI tables
> 
>> and
>> that is why there is the firmware which provides the information about
>> the hardware to the OS in a generic way.
>>
>>> Frankly with platform specific interfaces like ACPI, virtio-iommu
is
>>> much less compelling.  Describing topology as part of the device in
a
>>> way that is first, portable, and second, is a good fit for
hypervisors,
>>> is to me one of the main reasons virtio-iommu makes sense at all.
>>
>> Virtio-IOMMU makes sense in the first place because it is much faster
>> than emulating one of the hardware IOMMUs.
> 
> I don't see why it would be much faster. The interface isn't that
> different from command queues of VTD.
> 
>> And an ACPI table is also
>> portable to all ACPI platforms, same with device-tree.
>>
>> Regards,
>>
>> 	Joerg
> 
> Making ACPI meet the goals of embedded projects such as kata containers
> would be a gigantic task with huge stability implications.  By
> comparison this 400-line parser is well contained and does the job.  I
> didn't yet see compelling reasons not to merge this, but I'll be
> interested to see some more specific concerns.
> 
> 
Thanks

Eric

Hi Michael,

On Tue, Mar 03, 2020 at 11:09:41AM -0500, Michael S. Tsirkin
wrote:
> No. It's coded into the hardware. Which might even be practical
> for bare-metal (e.g. on-board flash), but is very practical
> when the device is part of a hypervisor.
If its that way on PPC, than fine for them. But since this is enablement
for x86, it should follow the x86 platform best practices, and that
means describing hardware through ACPI.
> This "hardware" is actually part of hypervisor so there's no
> reason it can't be completely self-descriptive. It's specified
> by the same entity as the "firmware".
That is just an implementation detail. Yes, QEMU emulates the hardware
and builds the ACPI tables. But it could also be implemented in a way
where the ACPI tables are build by guest firmware.
> I don't see why it would be much faster. The interface isn't that
> different from command queues of VTD.
VirtIO IOMMU doesn't need to build page-tables that the hypervisor then
has to shadow, which makes things much faster. If you emulate one of the
other IOMMUs (VT-d or AMD-Vi) the code has to shadow the full page-table
at once when device passthrough is used. VirtIO-IOMMU doesn't need that,
and that makes it much faster and efficient.
> Making ACPI meet the goals of embedded projects such as kata containers
> would be a gigantic task with huge stability implications.  By
> comparison this 400-line parser is well contained and does the job.  I
> didn't yet see compelling reasons not to merge this, but I'll be
> interested to see some more specific concerns.
An ACPI table parse wouldn't need more lines of code. For embedded
systems there is still the DT way of describing things.

Regards,

	Joerg

On Wed, Mar 04, 2020 at 07:48:54AM -0800, Jacob Pan
wrote:
> For emulated VT-d IOMMU, GIOVA can also be build as first level page
> tables then pass to the host IOMMU to bind. There is no need to shadow
> in this case, pIOMMU will do nested translation and walk guest page
> tables.
Right, but that requires hardware support. A pure software emulation of
VT-d requires the full shadow of the guest io-page-table.
> I thought we have the universal device properties to abstract DT and
> ACPI (via _DSD). Is that an option?
I don't know whether this was considered, Jean-Philippe?


Regards,

	Joerg