[libvirt-users] multiple devices in the same iommu group in L1 guest

Mon Aug 6 19:28:21 UTC 2018

On 07/03/2018 09:37 AM, Yalan Zhang wrote:
> Hi,
> 
> I have a guest enabled vIOMMU, but on the guest there are several
> devices in the same iommu group.
> Could someone help to check if I missed something? 
> Thank you very much!
> 
> 1. guest xml:
> # virsh edit q
> ...
> <os>
>     <type arch='x86_64' machine='pc-q35-rhel7.5.0'>hvm</type>
>     <loader readonly='yes' secure='yes'
> type='pflash'>/usr/share/OVMF/OVMF_CODE.secboot.fd</loader>
>     <nvram>/var/lib/libvirt/qemu/nvram/q_VARS.fd</nvram>
>   </os>
> ...
>  <features>
>  ...
>     <ioapic driver='qemu'/>
>   </features>
>  <cpu mode='host-passthrough' check='none'>
>     <feature policy='require' name='vmx'/>
>   </cpu>
> ...
> <devices>
> ...
>  <controller type='pci' index='7' model='pcie-root-port'>
>       <model name='pcie-root-port'/>
>       <target chassis='7' port='0x15'/>
>       <address type='pci' domain='0x0000' bus='0x00' slot='0x02'
> function='0x5'/>
>     </controller>
> <interface type='network'>
>       <mac address='52:54:00:b9:ff:90'/>
>       <source network='default'/>
>       <model type='e1000'/>
>       <address type='pci' domain='0x0000' bus='0x07' slot='0x00'
> function='0x0'/>
>     </interface>
> <iommu model='intel'>
>       <driver intremap='on' caching_mode='on' iotlb='on'/>
>     </iommu>
> </devices>
> ...
> 2. guest has 'intel_iommu=on' enabled in kernel cmdline, then reboot guest
> 
> 3. log in guest to check:
> # dmesg  | grep -i DMAR
> [    0.000000] ACPI: DMAR 000000007d83f000 00050 (v01 BOCHS  BXPCDMAR
> 00000001 BXPC 00000001)
> [    0.000000] DMAR: IOMMU enabled
> [    0.155178] DMAR: Host address width 39
> [    0.155180] DMAR: DRHD base: 0x000000fed90000 flags: 0x1
> [    0.155221] DMAR: dmar0: reg_base_addr fed90000 ver 1:0 cap
> 12008c22260286 ecap f00f5e
> [    0.155228] DMAR: ATSR flags: 0x1
> [    0.155231] DMAR-IR: IOAPIC id 0 under DRHD base  0xfed90000 IOMMU 0
> [    0.155232] DMAR-IR: Queued invalidation will be enabled to support
> x2apic and Intr-remapping.
> [    0.156843] DMAR-IR: Enabled IRQ remapping in x2apic mode
> [    2.112369] DMAR: No RMRR found
> [    2.112505] DMAR: dmar0: Using Queued invalidation
> [    2.112669] DMAR: Setting RMRR:
> [    2.112671] DMAR: Prepare 0-16MiB unity mapping for LPC
> [    2.112820] DMAR: Setting identity map for device 0000:00:1f.0 [0x0 -
> 0xffffff]
> [    2.211577] DMAR: Intel(R) Virtualization Technology for Directed I/O
> ===> This is expected
> 
> # dmesg  | grep -i iommu  |grep device
> [    2.212267] iommu: Adding device 0000:00:00.0 to group 0
> [    2.212287] iommu: Adding device 0000:00:01.0 to group 1
> [    2.212372] iommu: Adding device 0000:00:02.0 to group 2
> [    2.212392] iommu: Adding device 0000:00:02.1 to group 2
> [    2.212411] iommu: Adding device 0000:00:02.2 to group 2
> [    2.212444] iommu: Adding device 0000:00:02.3 to group 2
> [    2.212464] iommu: Adding device 0000:00:02.4 to group 2
> [    2.212482] iommu: Adding device 0000:00:02.5 to group 2
> [    2.212520] iommu: Adding device 0000:00:1d.0 to group 3
> [    2.212533] iommu: Adding device 0000:00:1d.1 to group 3
> [    2.212541] iommu: Adding device 0000:00:1d.2 to group 3
> [    2.212550] iommu: Adding device 0000:00:1d.7 to group 3
> [    2.212567] iommu: Adding device 0000:00:1f.0 to group 4
> [    2.212576] iommu: Adding device 0000:00:1f.2 to group 4
> [    2.212585] iommu: Adding device 0000:00:1f.3 to group 4
> [    2.212599] iommu: Adding device 0000:01:00.0 to group 2
> [    2.212605] iommu: Adding device 0000:02:01.0 to group 2
> [    2.212621] iommu: Adding device 0000:04:00.0 to group 2
> [    2.212634] iommu: Adding device 0000:05:00.0 to group 2
> [    2.212646] iommu: Adding device 0000:06:00.0 to group 2
> [    2.212657] iommu: Adding device 0000:07:00.0 to group 2
> ====> several devices in the same iommu group
> 
> # virsh nodedev-dumpxml pci_0000_07_00_0
> <device>
>   <name>pci_0000_07_00_0</name>
>   <path>/sys/devices/pci0000:00/0000:00:02.5/0000:07:00.0</path>
>   <parent>pci_0000_00_02_5</parent>
>   <driver>
>     <name>e1000</name>
>   </driver>
>   <capability type='pci'>
>     <domain>0</domain>
>     <bus>7</bus>
>     <slot>0</slot>
>     <function>0</function>
>     <product id='0x100e'>82540EM Gigabit Ethernet Controller</product>
>     <vendor id='0x8086'>Intel Corporation</vendor>
>     <iommuGroup number='2'>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x0'/>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x1'/>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x2'/>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x3'/>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x4'/>
>       <address domain='0x0000' bus='0x00' slot='0x02' function='0x5'/>
>       <address domain='0x0000' bus='0x01' slot='0x00' function='0x0'/>
>       <address domain='0x0000' bus='0x02' slot='0x01' function='0x0'/>
>       <address domain='0x0000' bus='0x04' slot='0x00' function='0x0'/>
>       <address domain='0x0000' bus='0x05' slot='0x00' function='0x0'/>
>       <address domain='0x0000' bus='0x06' slot='0x00' function='0x0'/>
>       <address domain='0x0000' bus='0x07' slot='0x00' function='0x0'/>
>     </iommuGroup>

All devices plugged into the same slot on pcie-root are in the same
iommu group, and all devices are in the same iommu group as the
pcie-root-port they are plugged into. Since libvirt groups 8
pcie-root-ports onto a single pcie-root slot (to conserve the relatively
limited number of slots), all the devices plugged into those 8
pcie-root-ports are in the same iommu group.

This can be solved by putting the pcie-root-ports *of the devices you
want to assign to L2 guests* onto pcie-root slots by themselves, i.e.
instead of addressing it at 00:02.5, address it at 00:10.0 (and don't
put any other pcie-root-ports at 00:10.x).

I had thought there was an open BZ about this, but can't find one
anywhere. At first glance it would seem that, since the pcie-root-ports
and iommu controller are all emulated devices, it would be simple to
just report all the pcie-root-ports as being on different iommu groups
regardless of their placement on pcie-root, but I believe when this was
brought up doubt was expressed over whether or not qemu truly can
guarantee that the devices are fully isolated from each other.

>   </capability>
> </device>
> 
> Thus, can not attach the device to L2 guest:
> # cat hostdev.xml
> <hostdev mode='subsystem' type='pci' managed='yes'>
>     <source>
>       <address domain='0x0000' bus='0x07' slot='0x00' function='0x0'/>
>     </source>
>   </hostdev>
> # virsh attach-device rhel hostdev.xml
> error: Failed to attach device from hostdev.xml
> error: internal error: unable to execute QEMU command 'device_add': vfio
> error: 0000:07:00.0: group 2 is not viable
> 
> 
> -------
> Best Regards,
> Yalan Zhang
> IRC: yalzhang
> Internal phone: 8389413
> 
> 
> _______________________________________________
> libvirt-users mailing list
> libvirt-users at redhat.com
> https://www.redhat.com/mailman/listinfo/libvirt-users
> 