[libvirt-users] PCI-Passthrough suddenly stopped working
Hi everybody! On a Centos 5.8 with libvirt qemu-0.10.5-1.el5.2, libvirt-0.8.2-22.el5 and 2.6.18-308.11.1.el5 I have a Windows machine that used to have pci-passthrough for a device enabled and working. After an update (kernel, libvirt afaik) and a reboot, pci-passthrough stopped working. Intel IOMMU is enabled (can see that from the dmesg output). All i get on the console when i try to start the vm is the following: Fehler: internal error process exited while connecting to monitor: char device redirected to /dev/pts/3 Failed to assign device "14:00.2" : Invalid argument Failed to deassign device "14:00.2" : Invalid argument Failed to initialize assigned device host=14:00.2 SELinux ist disabled. Does anyone have any hint what could have possibly gone wrong? As an additional info: "virsh nodeddev-list" also doesn't work. I don't know if I mix that up, but i seem to remember that that used to work on that machine. root@kvm ~]# virsh nodedev-list --tree Fehler: Failed to count node devices Fehler: this function is not supported by the connection driver: virNodeNumOfDevices Any hints are greatly appreciated! Thanks & Cheers, Rouven -- Blinkenlichten Open Source Solutions Maass Sacha GbR | Weigandufer 45 | 12059 Berlin tel: +493013896247 | fax: +493013896249 | mob: +491744220127 Web: http://www.blinkenlichten.de/ G+: http://gplus.to/blinkenlichten Blinkenlichten Zarafa Hosted Tweets: http://twitter.com/zarafamail/
On 07/11/2012 04:24 PM, Rouven Sacha wrote:
Hi everybody!
On a Centos 5.8 with libvirt qemu-0.10.5-1.el5.2, libvirt-0.8.2-22.el5 and 2.6.18-308.11.1.el5 I have a Windows machine that used to have pci-passthrough for a device enabled and working.
After an update (kernel, libvirt afaik) and a reboot, pci-passthrough stopped working. Intel IOMMU is enabled (can see that from the dmesg output).
This may be the result of a security fix in the new kernel. I know at least one older version of Intel chips has a bug where IOMMU can be exploited by a guest to take control over the host, so on those chips, newer kernels now require to explicitly enable a kernel module parameter to state that you are going to allow passthrough to the guest in spite of the security risk. That is, you may need to use: modprobe kvm allow_unsafe_assigned_interrupts=1 with your newer kernel. Unfortunately, I wasn't able to find a better URL to a page documenting this issue, so that implies we probably also need a patch to the libvirt documentation with regards to using device passthrough. -- Eric Blake eblake@redhat.com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
Am 12.07.2012 00:38, schrieb Eric Blake:
This may be the result of a security fix in the new kernel. I know at least one older version of Intel chips has a bug where IOMMU can be exploited by a guest to take control over the host, so on those chips, newer kernels now require to explicitly enable a kernel module parameter to state that you are going to allow passthrough to the guest in spite of the security risk. That is, you may need to use: modprobe kvm allow_unsafe_assigned_interrupts=1 with your newer kernel. Unfortunately, I wasn't able to find a better URL to a page documenting this issue, so that implies we probably also need a patch to the libvirt documentation with regards to using device passthrough.
Hi Eric, thanks for the info. Reading https://bugzilla.redhat.com/show_bug.cgi?id=715555 , it seems that 5.8 shouldn't be affected since the kvm on that version doesn't support interrupt remapping, if I understand correctly. Additionally, if I run the script provided in the issue description, the check passes with "Interrupt remapping support available" and the error message differs: I don't get "Operation not permitted" but "Invalid argument". I also can't set provides switch in /sys/module/kvm/parameters/allow_unsafe_assigned_interrupts, since the file isn't there on my box. Are there any other circumstances where pci passthrough could fail? Googling for the error message i get, I can't seem to find any case that matches mine. This makes me guess that I'd rather accidentally introduced a misconfiguration than encountered a qemu-kvm/libvirt bug. I have attached the configuration file of that machine, maybe someone could have a look at the hostdev section? Unfortunately, i am currently not able to switch back to the prior kernel, since the system is in production right now - I will test that later during the day. Thanks & cheers, Rouven -- Blinkenlichten Open Source Solutions Maass Sacha GbR | Weigandufer 45 | 12059 Berlin tel: +493013896247 | fax: +493013896249 | mob: +491744220127 Web: http://www.blinkenlichten.de/ G+: http://gplus.to/blinkenlichten Blinkenlichten Zarafa Hosted Tweets: http://twitter.com/zarafamail/
participants (2)
-
Eric Blake -
Rouven Sacha