[Kimchi-devel] [PATCH v7 2/4] Host device passthrough: List eligible device to passthrough

[Aline Manera]

On 07/13/2014 11:44 PM, Zhou Zheng Sheng wrote:
> on 2014/07/13 22:43, Aline Manera wrote:
>> On 07/09/2014 04:01 AM, Zhou Zheng Sheng wrote:
>>> This patch adds a '_passthrough=true' filter to /host/devices, so it can
>>> filter and shows all devices eligible to passthrough to guest.
>>> Theoretically, all PCI, USB and SCSI devices can be assigned to guest
>>> directly. However usually all host devices form a tree, if we assign a
>>> PCI port/SCSI controller/USB controller to guest, all devices/disks
>>> under the controller are assigned as well. In this patch we only present
>>> the "leaf" host devices to the user as potential passthrough devices.
>>> In other word, the possible devices are wireless network interface, SD
>>> card reader, camera, SCSI unit(disk or CD), and so on.
>>>
>>> Linux kernel is able to recognize the host IOMMU group layout. If two
>>> PCI devices are in the same IOMMU group, it means there are possible
>>> interconnections between the devices, and the devices can talk to each
>>> other bypassing IOMMU. This implies isolation is not pefect between those
>>> devices, so all devices in a IOMMU group must be assigned to guest
>>> together. On host that recognizes IOMMU groups, by accessing the URI
>>> /host/devices/deviceX/passthrough_affected_devices, it returns a list
>>> containing the devices in the same IOMMU group as deviceX.
>>>
>>> How to test:
>>>
>>> List all types of devices to passthrough
>>>     curl -k -u root -H "Content-Type: application/json" \
>>>       -H "Accept: application/json" \
>>>       'https://127.0.0.1:8001/host/devices?_passthrough=true'
>>>
>>> List all eligible PCI devices to passthrough
>>>     /host/devices?_passthrough=true&_cap=pci
>>>
>>> List all USB devices to passthrough
>>>     /host/devices?_passthrough=true&_cap=usb_device
>>>
>>> List all SCSI devices to passthrough
>>>     /host/devices?_passthrough=true&_cap=scsi
>>>
>>> List devices in the same IOMMU group as pci_0000_00_19_0
>>>     /host/devices/pci_0000_00_19_0/passthrough_affected_devices
>>
>> This passthrough_affected_devices seems a filter instead of a new resource.
>>
>> Couldn't we do: /host/devices/pci_0000_00_19_0?_passthrough=true ?
>>
>
> Thanks Aline. Let me explain why I don't think it's a filter.
>
> /host/devices/pci_0000_00_19_0 is a particular device, it contains
> information like vendor and product id/name, device sysfs patch, domain,
> bus, slot function id. All these are simple attributes of this device.
> The affected devices are a collection containing all devices in the same
> iommu group as pci_0000_00_19_0. If you think a new sub-collection is
> not good. I propose using a filter for /host/devices. For example,
>
> /host/devices?_passthrough_group_by=pci_0000_00_19_0

Thanks for the explanation, Zheng Sheng!

 From my view, your proposal using filter is better than a new 
sub-collection:

/host/devices?_passthrough_group_by=pci_0000_00_19_0


>
>>>
>>> v1:
>>>     v1 series does not contain this patch.
>>>
>>> v2:
>>>     Deal with calculation "leaf" device and "affected" device.
>>>
>>> v3 v4:
>>>     No change.
>>>
>>> v5:
>>>     Change _passthrough=1 to _passthrough=true in the URI scheme. Filter
>>> PCI devices according the PCI class.
>>>
>>> v6:
>>>     Don't passthrough PCI device of class code 07. In modern
>>> x86 machine, it's possible that
>>> "6 Series/C200 Series Chipset Family MEI Controller" and
>>> "6 Series/C200 Series Chipset Family KT Controller"
>>> are of this class code. These two devices are not suitable to
>>> passthrough to guest. We don't have simple and reliable way to
>>> distinguish normal serial controller and host chipset XXX controller.
>>> This type of PCI devices also include various serial, parallel, modem,
>>> communication controller. Serial and parallel controllers can be
>>> re-direct from ttyS0 to QEMU's pty using socat, and there is little
>>> performance benefit to directly assign to guest. So it'k ok not to
>>> passththrough PCI device of class code 07.
>>>
>>> Signed-off-by: Zhou Zheng Sheng <zhshzhou at linux.vnet.ibm.com>
>
> Didn't find more comments below, so I truncate the mail.
>