I upgraded to a shiny new AMD Ryzen 3rd gen processer (Threadripper 3960x.) After doing so I could not boot up my Windows 10 gaming VM (it uses VFIO / PCI Passthrough for the video card.) The message I kept getting as it tried to boot was:
The problem comes with a new speculative execution protection hardware feature in the Ryzen Gen 3 chipsets – stibp. Qemu doesn’t know how to handle it properly, thus the bluescreens.
There are two ways to fix it
Change host-model from host-passthrough to epyc
Add CPU parameters to your Virtual Machine’s XML file instructing it to not use the stibp CPU feature.
Since I have some software that checks CPU model and refuses to work if it’s not in the desktop class (Geforce Experience) I opted for route #2.
First, check the qemu logs to see which CPU parameters your VM was using (pick a time where it worked.) Replace ‘win10’ with the name of your VM.
in my case, it was -cpu host,migratable=on,topoext=on,kvmclock=on,hv-time,hv-relaxed,hv-vapic,hv-spinlocks=0x1fff,hv-vendor-id=1234567890ab,kvm=off \
Copy everything after -cpu and before the last backslash. Then edit your VM’s XML file (change last argument to the name of your VM)
sudo virsh edit win10
Scroll down to the bottom qemu:commandline section (if it doesn’t exist, create it right above the last line – </domain>. Paste the following information obtained from the above log (ignoring the qemu:commandline lines if they already exist.) In my case it looked like this:
What you’re doing is copying the CPU arguments you found in the log and adding them to the qemu:commandline section, with a twist – adding -amd-stibp which instructs qemu to remove that CPU flag.
I’ve once again switched from Proxmox to Arch Linux for my desktop machine. Both use KVM so it’s really just a matter of using the different VM manager syntax (virt-manager vs qm.) I used my notes from my previous stint with Arch, my article on GPU Passthrough in Proxmox as well as a thorough reading of the Arch wiki’s PCI Passthrough article.
Enable IOMMU
Configure GRUB to load the necessary iommu modules at boot. Append amd_iommu=on iommu=pt to the end of GRUB_CMDLINE_LINUX_DEFAULT (change accordingly if you have Intel instead of AMD)
sudo vim /etc/default/grub
...
GRUB_CMDLINE_LINUX_DEFAULT="loglevel=3 amd_iommu=on iommu=pt"
Run update-grub
sudo update-grub
Reserve GPU for VFIO
Reserve the GPU you wish to pass through to a VM for use with the vfio kernel driver (so the host OS doesn’t interfere with it)
Determine the PCI address of your GPU
Run lspci -v and look for your card. Mine was 01:00.0 & 01:00.1. You can omit the part after the decimal to include them both in one go – so in that case it would be 01:00
Run lspci -n -s <PCI address from above> to obtain vendor IDs. Example : lspci -n -s 01:00 01:00.0 0300: 10de:1b81 (rev a1) 01:00.1 0403: 10de:10f0 (rev a1)
Assign your GPU to vfio driver using the IDs obtained above. Example using above IDs: echo "options vfio-pci ids=10de:1b81,10de:10f0" >> /etc/modprobe.d/vfio.conf
Reboot the host to put the kernel / drivers into effect.
Assuming you’re using network manager for your connections, create a bridge (thanks to ciberciti.biz & the arch wiki for information on how to do so.) Replace interface names with ones corresponding to your machine:
sudo nmcli connection add type bridge ifname br0 stp no
sudo nmcli connection add type bridge-slave ifname enp4s0 master br0
sudo nmcli connection show
#Make note of the active connection name
sudo nmcli connection down "Wired connection 2" #from above
sudo nmcli connection up bridge-br0
Create a second bridge bound to lo0 for host-only communication. Change IP as desired:
sudo nmcli connection add type bridge ifname br99 stp no ip4 192.168.2.1/24
sudo nmcli connection add type bridge-slave ifname lo master br99
sudo nmcli connection up bridge-br99
Configure VM
Initial configuration
When creating the passthrough VM, make sure chipset is Q35.
Set the CPU model to host-passthrough (type it in, there is no dropdown for it.)
When adding disks / other devices, set the device model to virtio
Add your GPU by going to Add Hardware and finding it under PCI Host Device.
Windows 10 specific tweaks
If your passthrough VM is going to be windows based, some tweaks are required to get the GPU to work properly within the VM.
Ignore MSRs (blue screen fix)
Later versions of Windows 10 instantly bluescreen with kmode_exception_not_handled unless you pass an option to ignore MSRs. Add the kvm ignore_msrs=1 option in /etc/modprobe.d/kvm.conf to do so. Optionally add the report_ignored_msrs=0 option to squelch massive amounts of kernel messages every time an MSR was ignored.
Use the virsh edit command to make some tweaks to the VM configuration. We need to hide the fact that this is a VM otherwise the GPU drivers will not load and will throw Error 43. We need to add a vendor_id in the hyperv section, and create a kvm section enabling hidden state, which hides certain CPU flags that the drivers use to detect if they’re in a VM or not.
If you operate on a multi-core system such as my AMD Ryzen Threadripper the you will want to optimize your CPU core configuration in the VM per the CPU Pinning section in the Arch Wiki
Determine your CPU topology by running lscpu -e and lstopo The important things to look for are the CPU number and core number. On my box, it looks like this:
From the above output I see my CPU core 0 is shared by CPUs 0 & 16, meaning CPU 0 and CPU 16 (as seen by the Linux kernel) are hyperthreaded to the same physical CPU core.
Especially for gaming, you want to keep all threads on the same CPU cores (for multithreading) and the same CPU die (on my threadripper, CPUs 0-7 reside on one physical die, and CPUs 8-15 reside on the other, within the same socket.)
In my case I want to dedicate one CPU die to my VM with its accompanying hyperthreads (CPUs 0-7 & hyperthreads 16-23) You can accomplish this using the virsh edit command and creating a cputune section (make sure you have a matching vcpu count for the number of cores you’re configuring.) Also edit CPU mode with the proper topology of 1 socket, 1 die, 8 cores with 2 threads. Lastly, configure memory to only be from the proper NUMA node the CPU cores your VM is using (Read here for more info.)
Non-uniform memory access is essential for 1st and 2nd gen Ryzen chips. It turns out that by default my motherboard hid the real NUMA configuration from the operating system. Remedy this by changing the BIOS setting to set Memory Interleaving = Channel (for my ASRock X399 motherboard it’s in CBS / DF options.) See here: https://www.reddit.com/r/Amd/comments/6vrcq0/psa_threadripper_umanuma_setting_in_bios/
After changing BIOS setting, lstopo now shows proper configuration:
CPU frequency
Change CPU frequency setting to use performance mode:
Append default_hugepagesz=1G hugepagesz=1G hugepages=16 to the kernel line in /etc/default/grub and re-run sudo grub-mkconfig -o /boot/grub/grub.cfg
Configure FIFO CPU scheduling
The Arch Wiki mentions to run qemu-system-x86_64 with taskset and chrt but doesn’t mention how to do so if you’re using virt-manager. Fortunately this reddit thread outlined how to accomplish it: libvirt hooks. Create the following script and place it in /etc/libvirt/hooks/qemu , change the VM variable to match the name of your VM, mark that new file as executable (chmod +x /etc/libvirt/hooks/qemu ) and restart libvirtd
#!/bin/bash
#Hook to change VM to FIFO scheduling to decrease latency
#Place this file in /etc/libvirt/hooks/qemu and mark it executable
#Change the VM variable to match the name of your VM
VM="win10"
if [ "$1" == "$VM" ] && [ "$2" == "started" ]; then
if pid=$(pidof qemu-system-x86_64); then
chrt -f -p 1 $pid
echo $(date) changing CPU scheduling to FIFO for VM $1 pid $pid >> /var/log/libvirthook.log
else
echo $(date) Unable to acquire PID of $1 >> /var/log/libvirthook.log
fi
fi
#Additional debug
#echo $(date) libvirt hook arg1=$1 arg2=$2 arg3=$3 arg4=$4 pid=$pid >> /var/log/libvirthook.log
Isolate CPUs
Update 7/28/20: I no longer do this in favor of the qemu hook script above, which prioritizes to p1 the qemu process for the cores it needs. I’m leaving this section here for historical/additional tweaking purposes.
Update 6/28/20: Additional tuning since I was having some stuttering and framerate issues. Also read here about the emulatorpin option
Dedicate CPUs to the VM (host will not use them) – append isolcups, nohz_full & rcu_nocbs kernel parameters into /etc/default/grub
taskset -cp 1
pid 1's current affinity list: 8-15,24-31
You can still tell programs to use the CPUs the VM has manually with the taskset command:
chrt -r 1 taskset -c <cores to use> <name of program/process>
Low Latency Audio
Upbate 7/8/2020: I found this article and this reddit thread (and this one) on how to use pulseaudio for your guest VM to get low latency guest VM audio piped to the host machine.
Update qemu config
edit /etc/libvirt/qemu.conf: uncomment the line #user = "root" and replace “root” with your username
Update pulseaudio config
Edit /etc/pulse/daemon.conf and uncomment the following lines (remove semicolon)
I had a pet project (folding@home) where I wanted to maximize computing power. I became frustrated with default CPU scheduling of my folding@home threads. Ideal performance would keep similar threads on the same CPU, but the threads were jumping all over the place, which was impacting performance.
Step one was to figure out which threads belonged to which physical cores. I found on this site that you can use cat to find out what your “sibling threads” are:
The above command is for my Threadripper & Epyc systems, which each have 16 cores hyperthreaded to 32 cores. Adjust the {0..15} number to match your number of cores (core 0 being the fist core.) This was my output:
Now that I know the sibling threads are offset by 16, I can use this information to optimize my folding@home VMs. I modified my CPU pinning script to take this into consideration. The script ensures that each VM is pinned to only use sibling threads (ensuring they all stay on the same physical CPU.)
This script should be used with caution. It pins processes to specific CPUs, which limits the kernel scheduler’s ability to move things around if needed. If configured badly this can cause the machine to lock up or VMs to be terminated.
I saw some impressive results spinning up four separate 8 core VMs and pinning them to sibling cores using this script. It almost doubled the rate at which I completed folding@home work units.
And now, the script:
#!/bin/bash
#Properly assign CPU cores to their respective die for EPYC/Threadripper systems
#Based on how hyperthreads are done in these systems
#cat /sys/devices/system/cpu/cpu{0..15}/topology/thread_siblings_list
#The script takes two arguments - the ID of the Proxmox VM to modify, and the core to begin the VM on
#If running this against multiple VMs, make sure to increment this second number by half of the cores of the previous VM
#For example, if I have one 8 core VM and I run this script specifying 0 for the offset, if I spin up a second VM, the second argument would be 4
#this would ensure the second VM starts on core 4 (the 5th core) and assigns sibling cores to match
set -eo pipefail
#take First argument as which VMID to pin CPU cores to, the second argument is which core to start pinning to
VMID=$1
OFFSET=$2
#Determine offset for sibling threads
SIBLING_THREAD_OFFSET=$(cat /sys/devices/system/cpu/cpu0/topology/thread_siblings_list| sed 's/,/ /g' | awk '{print $2}')
#Function to determine number of CPU cores a VM has
cpu_tasks() {
expect <<EOF | sed -n 's/^.* CPU .*thread_id=\(.*\)$/\1/p' | tr -d '\r' || true
spawn qm monitor $VMID
expect ">"
send "info cpus\r"
expect ">"
EOF
}
#Only act if VMID & OFFSET are set
if [[ -z $VMID || -z $OFFSET ]]
then
echo "Usage: cpupin.sh <VMID> <OFFSET>"
exit 1
else
#Get PIDs of each CPU core for VM, count number of VM cores, and get even/odd PIDs for assignment
VCPUS=($(cpu_tasks))
VCPU_COUNT="${#VCPUS[@]}"
VCPU_EVEN_THREADS=($(for EVEN_THREAD in "${VCPUS[@]}"; do echo $EVEN_THREAD; done | awk '!(NR%2)'))
VCPU_ODD_THREADS=($(for ODD_THREAD in "${VCPUS[@]}"; do echo $ODD_THREAD; done | awk '(NR%2)'))
if [[ $VCPU_COUNT -eq 0 ]]; then
echo "* No VCPUS for VM$VMID"
exit 1
fi
echo "* Detected ${#VCPUS[@]} assigned to VM$VMID..."
echo "* Resetting cpu shield..."
#Start at offset CPU number, assign odd numbered PIDs to their own CPU thread, then increment CPU core number
#0-3 if offset is 0, 4-7 if offset is 4, etc
ODD_CPU_INDEX=$OFFSET
for PID in "${VCPU_ODD_THREADS[@]}"
do
echo "* Assigning ODD thread $ODD_CPU_INDEX to $PID..."
taskset -pc "$ODD_CPU_INDEX" "$PID"
((ODD_CPU_INDEX+=1))
done
#Start at offset + CPU count, assign even number PIDs to their own CPU thread, then increment CPU core number
#16-19 if offset is 0, 20-23 if offset is 4, etc
EVEN_CPU_INDEX=$(($OFFSET + $SIBLING_THREAD_OFFSET))
for PID in "${VCPU_EVEN_THREADS[@]}"
do
echo "* Assigning EVEN thread $EVEN_CPU_INDEX to $PID..."
taskset -pc "$EVEN_CPU_INDEX" "$PID"
((EVEN_CPU_INDEX+=1))
done
fi
An unfortunate bug exists for the AMD Threadripper family of GPUs which causes VGA Passthrough not to work properly. Fortunately some very clever people have implemented a workaround to allow proper VGA passthrough until a proper Linux Kernel patch can be accepted and implemented. See here for the whole story.
Right now my Thrdearipper 1950x successfully has GPU passthrough thanks to HyenaCheeseHeads “java hack” applet. I went this route because I really didn’t want to try and recompile my ProxMox kernel to get passthrough to work. Per the description “It is a small program that runs as any user with read/write access to sysfs (this small guide assumes “root”). The program monitors any PCIe device that is connected to VFIO-PCI when the program starts, if the device disconnects due to the issues described in this post then the program tries to re-connect the device by rewriting the bridge configuration.” Instructions taken from the above Reddit post.
Go to https://pastebin.com/iYg3Dngs and hit “Download” (the MD5 sum is supposed to be 91914b021b890d778f4055bcc5f41002)
Rename the downloaded file to “ZenBridgeBaconRecovery.java” and put it in a new folder somewhere
Go to the folder in a terminal and type “javac ZenBridgeBaconRecovery.java”, this should take a short while and then complete with no errors. You may need to install the Java 8 JDK to get the javac command (use your distribution’s software manager)
In the same folder type “sudo java ZenBridgeBaconRecovery”
Make sure that the PCIe device that you intend to passthru is listed as monitored with a bridge
Now start your VM
In my case (Debian Stretch, ProxMox) I needed to install openjdk-8-jdk-headless
Next I have a little script on startup to spawn this as root in a detached tmux session, so I don’t have to remember to run it (If you try to start your VM before running this, it will hose passthrough on your system until you reboot it.) Be sure to change the script to point to wherever you compiled ZenBridgeBaconRecovery
#!/bin/bash
cd /home/nicholas #change me to suit your needs
sudo java ZenBridgeBaconRecovery
And here is the command I use to run on startup:
tmux new -d '/home/nicholas/passthrough.sh'
Again, be sure to modify the above to point to the path of wherever you saved the above script.
So far this works pretty well for me. I hate having to run a java process as sudo, but it’s better than recompiling my kernel.
Update 6/27/2018: I’ve created a systemd service script for the ZenBaconRecovery file to run at boot. Here is my file, placed in
/etc/systemd/system/zenbridge.service: (change your working directory to match the zenbridgebaconrecovery java file location. Don’t forget to do systemctl daemon-reload.)
Per an update on the reddit thread motherboard manufactures have finally put out BIOS updates that resolve the PCI passthrough problems. I updated my X399 Tachi to the latest version of its UEFI BIOS (3.20) and indeed PCI passthrough worked without any more wonky workarounds!
