Docker-OSX/helm/INSTALL-QEMU-AND-GPU-IOMMU.md

Install macOS Docker Virtualization

Setup

This walks through setting up QEMU virtualization for running macOS in Docker & Kubernetes

Please note, this guide assumes the host operation system is running Centos 7 (or ClearOS 7 more specifically). These commands can mostly be transferred to other distros, but there are a few areas that need commands (i.e. updating )

Host configuration

Build QEMU and libvirt from source

Since there is no official QEMU 5.X repo it appears, build from source.

QEMU Requirements

Python, glib2-devel, and pixman

sudo yum install python glib2-devel cairo-devel -y

Ninja

pip3 install ninja

Build QEMU from source steps

Clone the offical QEMU repo and build from source:

git clone git://git.qemu-project.org/qemu.git
cd qemu
mkdir -p bin/debug/native
cd bin/debug/native
../../../configure --enable-debug
make -j24
make install

Note: adjust make to use the desired number of threads avaliable on your system

libvirt Requirements

Configure repo:

yum-config-manager --nogpgcheck --add-repo http://mirror.centos.org/centos/7/virt/x86_64/libvirt-latest/

Install libvirt

yum install libvirt -y

Update permissions

chmod 660 -R /dev/kvm && chown 1000:1000 /dev/kvm
usermod -a -G kvm root

Note: these may not be required

Verification

Ensure latest version installed

virsh -c qemu:///system version --daemon

• For example, should output something like:

  [root@server repos]# virsh -c qemu:///system version --daemon
  Compiled against library: libvirt 5.0.0
  Using library: libvirt 5.0.0
  Using API: QEMU 5.0.0
  Running hypervisor: QEMU 5.2.50
  Running against daemon: 5.0.0
  

Install IMMO for GPU passthrough

1. Modify GRUB boot args:

   Add the following to /etc/default/grub to the end of the GRUB_CMDLINE_LINUX parameter:

   GRUB_CMDLINE_LINUX="... iommu=pt intel_iommu=on"
   

2. Update GRUB2:

   grub2-mkconfig -o /boot/efi/EFI/clearos/grub.cfg
   

   Note: this command may vary based on location of the grub.cfg for the boot entry

3. Reboot system

4. Ensure that the kernel parameter changes worked:

   cat /proc/cmdline
   

5. Find GPU hardware ids with lspci

   Example:

   lspci -nn | grep -i nvidia
   

6. Add the hardware ids to /etc/modprobe.d/vfio.conf

   Example:

   options vfio-pci ids=10de:1b81,10de:10f0
   

   Note: this is for the NVIDIA GTX 1070

7. Enable vfio-pci

   echo 'vfio-pci' > /etc/modules-load.d/vfio-pci.conf
   

   Make backup and rebuild initramfs:

   cp -p /boot/initramfs-$(uname -r).img /boot/initramfs-$(uname -r).img.bak
   dracut -f 
   

   Note: dracut -f may take awhile..

8. Increase ulimits

   This is done to avoid memory issues like VFIO_MAP_DMA: -12 and etc

   Append the following to /etc/security/limits.conf:

   @kvm            soft    memlock         unlimited
   @kvm            hard    memlock         unlimited
   

   Append the following to /etc/docker/daemon.json:

   {
       "default-ulimits": {
           "nofile": {
               "name": "nofile",
               "hard": 65536,
               "soft": 1024
           },
           "memlock":
           {
               "name": "memlock",
               "soft": -1,
               "hard": -1
           }
       }
   }
   

   Add LimitMEMLOCK to /etc/systemd/system/multi-user.target.wants/libvirtd.service like:

   [Unit]
   Description=Virtualization daemon
   ...
   
   [Service]
   ...
   LimitMEMLOCK=infinity
   

9. Reload systemd after changing config

   systemctl daemon-reload
   

10. Reboot system

11. Ensure that vfio worked

    dmesg | grep -i vfio
    

Issues

Many issues can rise up as a result of adding the complexity layers involved here. Some of the main areas are improperly loading the vfio-pci driver for the GPU and permission issues.

Modules for vfio not loading

When vfio does not load, errors such as the following can be seen:

error getting device from group *: No such device
Verify all devices in group * are bound to vfio-<bus> or pci-stub and not already in use

This can show up when vfio-pci driver is not loaded for the peripheral. Ensure that vfio-pci is loaded.

dmesg | grep -i vfio

If so, explicitly tell vfio modules to start

echo 'vfio
vfio_iommu_type1
vfio_pci
vfio_virqfd' > /etc/modules

Make backup and rebuild initramfs:

cp -p /boot/initramfs-$(uname -r).img /boot/initramfs-$(uname -r).img.bak
dracut -f 

Note: dracut -f may take awhile..

Do a system reboot

After rebooting, check on the gpu with lspci utilizing your gpu hardware id:

I.E.

[root@server docker-docker-osx]# lspci -vvv -s 09:00.0
09:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Ellesmere [Radeon RX 470/480/570/570X/580/580X/590] (rev c7) (prog-if 00 [VGA controller])
        Subsystem: Advanced Micro Devices, Inc. [AMD/ATI] Radeon RX 480
        Physical Slot: 5
        Control: I/O- Mem- BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr+ Stepping- SERR+ FastB2B- DisINTx-
        Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx-
        Interrupt: pin A routed to IRQ 255
        ...
        Kernel driver in use: vfio-pci
        Kernel modules: amdgpu

It does not matter if the host os loads a gpu module as seen with Kernel modules: amdgpu in the case above, the important part is that vfio-pci is the driver in use.

Permissions on vfio and kvm

One of the biggest areas of pain can be setting permissions on /dev/kvm, /dev/vfio/vfio, or /dev/vfio/<iommu_group>. If permission errors are seen, try the following commands:

chmod 660 -R /dev/kvm && chown 1000:1000 /dev/kvm
chmod 777 -R /dev/vfio && chown 1000:1000 -R /dev/vfio

References

https://gist.github.com/dghubble/c2dc319249b156db06aff1d49c15272e

Configure IOMMU and vfio https://www.server-world.info/en/note?os=CentOS_7&p=kvm&f=10

Configuring GPU driver with vfio-pci binding https://github.com/intel/nemu/wiki/Testing-VFIO-with-GPU

IOMMU Interrupt Mapping https://pve.proxmox.com/wiki/Pci_passthrough#IOMMU_Interrupt_Remapping

Manual Graphics Driver Binding https://lwn.net/Articles/143397/

QEMU Stdio Example https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg04521.html