Re: [libvirt-users] [Qemu-devel] VFIO PCIe Extended Capabilities
Hi Marcel,
Indeed, if a device is attached to a PCI bus it makes no sense to advertise the extended configuration space. Can you please share the QEMU command line? Maybe is possible to make the device's bus PCIe in QEMU?
Changing the following should make the tweak I proposed irrelevant. - device vfio-pci,host=03:00.0,id=hostdev0,bus=pci.2,addr=0x4 + device vfio-pci,host=03:00.0,id=hostdev0,bus=pcie.0,addr=0x4 Here's the full command for reference. 2016-07-19 18:42:22.110+0000: starting up libvirt version: 1.2.17, package: 13.el7 (CentOS BuildSystem <http://bugs.centos.org>, 2015-11-20-16:24:10, worker1.bsys.centos.org), qemu version: 2.3.0 (qemu-kvm-ev-2.3.0-31.el7_2.10.9) LC_ALL=C PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin QEMU_AUDIO_DRV=none /usr/libexec/qemu-kvm -name instance-00000019 -S -machine pc-q35-rhel7.2.0,accel=kvm,usb=off -cpu Haswell-noTSX,+abm,+pdpe1gb,+rdrand,+f16c,+osxsave,+dca,+pdcm,+xtpr,+tm2,+est,+smx,+vmx,+ds_cpl,+monitor,+dtes64,+pbe,+tm,+ht,+ss,+acpi,+ds,+vme -m 8192 -realtime mlock=off -smp 8,sockets=8,cores=1,threads=1 -uuid 8b6865eb-2118-430e-a7e0-2989696576b1 -smbios type=1,manufacturer=Fedora Project,product=OpenStack Nova,version=13.1.0-1.el7,serial=a1068a93-24d1-4da2-8903-f9b8307fb0d8,uuid=8b6865eb-2118-430e-a7e0-2989696576b1,family=Virtual Machine -no-user-config -nodefaults -chardev socket,id=charmonitor,path=/var/lib/libvirt/qemu/domain-instance-00000019/monitor.sock,server,nowait -mon chardev=charmonitor,id=monitor,mode=control -rtc base=utc,driftfix=slew -global kvm-pit.lost_tick_policy=discard -no-hpet -no-shutdown -boot strict=on -device i82801b11-bridge,id=pci.1,bus=pcie.0,addr=0x1e -device pci-bridge,chassis_nr=2,id=pci.2,bus=pci.1,addr=0x1 -drive file=/var/lib/nova/instances/8b6865eb-2118-430e-a7e0-2989696576b1/disk,if=none,id=drive-virtio-disk0,format=qcow2,cache=none -device virtio-blk-pci,scsi=off,bus=pci.2,addr=0x2,drive=drive-virtio-disk0,id=virtio-disk0,bootindex=1 -drive file=/var/lib/nova/instances/8b6865eb-2118-430e-a7e0-2989696576b1/disk.swap,if=none,id=drive-virtio-disk1,format=qcow2,cache=none -device virtio-blk-pci,scsi=off,bus=pci.2,addr=0x3,drive=drive-virtio-disk1,id=virtio-disk1 -netdev tap,fd=29,id=hostnet0,vhost=on,vhostfd=30 -device virtio-net-pci,netdev=hostnet0,id=net0,mac=fa:16:3e:9c:cf:60,bus=pci.2,addr=0x1 -chardev file,id=charserial0,path=/var/lib/nova/instances/8b6865eb-2118-430e-a7e0-2989696576b1/console.log -device isa-serial,chardev=charserial0,id=serial0 -chardev pty,id=charserial1 -device isa-serial,chardev=charserial1,id=serial1 -vnc 0.0.0.0:0 -k en-us -device cirrus-vga,id=video0,bus=pcie.0,addr=0x1 -device vfio-pci,host=03:00.0,id=hostdev0,bus=pci.2,addr=0x4 -device virtio-balloon-pci,id=balloon0,bus=pci.2,addr=0x5 -msg timestamp=on char device redirected to /dev/pts/3 (label charserial1) It seems very reasonable to change the qemu command line. However, I'm using OpenStack Nova to launch the instance, which then uses libvirt, which finally uses qemu. So, the issues is very buried.
I think that any instance of a q35 machine where the assigned device is placed on the conventional PCI bridge will create this scenario. It's the default for attaching devices to a libvirt managed q35 VM AFAIK.
Yes, I discussed with Laine from libvirt the possibility to assign devices to a PCIe port instead.
Yes, I see this as the issue as well. The VM is being created by libvirt with the following xml. The problem is that the device is auto-assigned to the pci-bridge by default. <domain type="kvm"> <uuid>8b6865eb-2118-430e-a7e0-2989696576b1</uuid> <name>instance-00000019</name> <memory>8388608</memory> <vcpu>8</vcpu> ... <snip> ... <devices> ... <snip> ... <hostdev mode="subsystem" type="pci" managed="yes"> <source> <address bus="0x03" domain="0x0000" function="0x0" slot="0x00"/> </source> </hostdev> ... <snip> ... </devices> </domain> If I do a dumpxml I get the following <domain type="kvm"> <uuid>8b6865eb-2118-430e-a7e0-2989696576b1</uuid> <name>instance-00000019</name> <memory>8388608</memory> <vcpu>8</vcpu> ... <snip> ... <devices> ... <snip> ... <controller type='pci' index='0' model='pcie-root'> <alias name='pcie.0'/> </controller> <controller type='pci' index='1' model='dmi-to-pci-bridge'> <alias name='pci.1'/> <address type='pci' domain='0x0000' bus='0x00' slot='0x1e' function='0x0'/> </controller> <controller type='pci' index='2' model='pci-bridge'> <alias name='pci.2'/> <address type='pci' domain='0x0000' bus='0x01' slot='0x01' function='0x0'/> </controller> <interface type='bridge'> ... <snip> ... <hostdev mode='subsystem' type='pci' managed='yes'> <driver name='vfio'/> <source> <address domain='0x0000' bus='0x03' slot='0x00' function='0x0'/> </source> <alias name='hostdev0'/> <address type='pci' domain='0x0000' bus='0x02' slot='0x02' function='0x0'/> </hostdev> </devices> </domain> If I manually change the domain as follows <hostdev mode='subsystem' type='pci' managed='yes'> <driver name='vfio'/> <source> <address domain='0x0000' bus='0x03' slot='0x00' function='0x0'/> </source> <alias name='hostdev0'/> - <address type='pci' domain='0x0000' bus='0x02' slot='0x02' function='0x0'/> + <address type='pci' domain='0x0000' bus='0x00' slot='0x02' function='0x0'/> </hostdev> I can reimport and the device and it attaches successfully to the pcie-root hub. The problem is I need to either manually specify this in nova or update the behavior of libvirt to auto assign pcie devices to pcie busses. This would involve reading pci configuration space to check if the device is a pcie devices. Then attaching to the pcie-root, if possible.
In fact, I've tried to fix this multiple times:
https://lists.nongnu.org/archive/html/qemu-devel/2015-10/msg05384.html https://lists.nongnu.org/archive/html/qemu-devel/2015-11/msg02422.html https://lists.nongnu.org/archive/html/qemu-devel/2016-01/msg03259.html
Yet the patch remains unapplied :(
I thought is it in already. Maybe Michael can add it as part of the hard freeze. And if the patch will be applied, the tweak above wouldn't help, right Alex?
The tweak Spenser suggested above should be unnecessary with my proposed patch applied.
I think I finally understand this. The bus is not pcie -> we return from vfio_add_ext_cap without "breaking" the extended capabilities chain and the bare metal SR-IOV capability will be visible to guest. With your patch the PCI bridge will "interfere" and mask the extended configuration space completely.
Only now searching for that patch did I notice Michael's comment hidden at the bottom of his reply, which I assume is why it never got applied:
I just saw it too! It seems Michael wants to cache this info in device instead of re-calculating it every time.
Anyway, the current behavior is clearly a bug, so QEMU hard freeze should be irrelevant. If anyone wants to take over the patch, feel free. Thanks,
I suppose I can handle it, but sadly not for 2.7. If Spencer has some time now he can help by testing it and reviewing it quickly :)
I'd be happy to help; but that patch really just more permanently breaks my current workaround ;-) . Thanks, Spenser This email and any attachments are intended for the sole use of the named recipient(s) and contain(s) confidential information that may be proprietary, privileged or copyrighted under applicable law. If you are not the intended recipient, do not read, copy, or forward this email message or any attachments. Delete this email message and any attachments immediately.
On Tue, 19 Jul 2016 20:39:26 +0000 Spenser Gilliland <spenser.gilliland@xilinx.com> wrote:
Anyway, the current behavior is clearly a bug, so QEMU hard freeze should be irrelevant. If anyone wants to take over the patch, feel free. Thanks,
I suppose I can handle it, but sadly not for 2.7. If Spencer has some time now he can help by testing it and reviewing it quickly :)
I'd be happy to help; but that patch really just more permanently breaks my current workaround ;-) .
How so? The patch should make extended config space disappear when the device is not on a PCIe bus. This is the correct behavior per the PCIe spec. Are you trying to exploit this QEMU bug to see some PCIe capabilities but not others? That's wrong, you'll need to move the device to a PCIe slot to get extended capabilities.
This email and any attachments are intended for the sole use of the named recipient(s) and contain(s) confidential information that may be proprietary, privileged or copyrighted under applicable law. If you are not the intended recipient, do not read, copy, or forward this email message or any attachments. Delete this email message and any attachments immediately.
You're sending email to a public list, please stop including these. Thanks, Alex
Hey Alex,
How so? The patch should make extended config space disappear when the device is not on a PCIe bus. This is the correct behavior per the PCIe spec. Are you trying to exploit this QEMU bug to see some PCIe capabilities but not others? That's wrong, you'll need to move the device to a PCIe slot to get extended capabilities.
Yes, I am exploiting the peculiarities of QEMU to make my hardware work properly. It may be wrong but it works (hence the ;-) ) .
From this conversation, it seems the real problem is that libvirt does not auto assign the hardware to a pcie root port in the q35 machine type. It seems like this really needs to be fixed in the libvirt source code; so, I'm working on a patch for libvirt.
I'd be happy to test your patch for you; I have access to hardware that will allow me to verify that your patch works correctly. Feel free to cc me on the next version of the patch and I'll add my tested-by.
You're sending email to a public list, please stop including these.
This is not intentional. It seems like the Xilinx email servers (not my local machine) are adding this to all messages. I'll see what I can do and I'm sorry for the inconvenience. Thanks, Spenser This email and any attachments are intended for the sole use of the named recipient(s) and contain(s) confidential information that may be proprietary, privileged or copyrighted under applicable law. If you are not the intended recipient, do not read, copy, or forward this email message or any attachments. Delete this email message and any attachments immediately.
participants (2)
-
Alex Williamson -
Spenser Gilliland