5.4. VFIO Advanced

5.4.1. VFIO no-IOMMU mode

If there is no IOMMU available on the system, VFIO can still be used, but it has to be loaded with an additional module parameter:

modprobe vfio enable_unsafe_noiommu_mode=1

Alternatively, one can also enable this option in an already loaded kernel module:

echo 1 > /sys/module/vfio/parameters/enable_unsafe_noiommu_mode

After that, VFIO can be used with hardware devices as usual.

Note

It may be required to unload all VFIO related-modules before probing the module again with enable_unsafe_noiommu_mode=1 parameter.

Warning

Since no-IOMMU mode forgoes IOMMU protection, it is inherently unsafe. That said, it does make it possible for the user to keep the degree of device access and programming that VFIO has, in situations where IOMMU is not available.

5.4.2. VFIO Memory Mapping Limits

For DMA mapping of either external memory or hugepages, VFIO interface is used. VFIO does not support partial unmap of once mapped memory. Hence DPDK’s memory is mapped in hugepage granularity or system page granularity. Number of DMA mappings is limited by kernel with user locked memory limit of a process (rlimit) for system/hugepage memory. Another per-container overall limit applicable both for external memory and system memory was added in kernel 5.1 defined by VFIO module parameter dma_entry_limit with a default value of 64K. When application is out of DMA entries, these limits need to be adjusted to increase the allowed limit.

5.4.3. Creating Virtual Functions using vfio-pci

Since Linux version 5.7, the vfio-pci module supports the creation of virtual functions. After the PF is bound to vfio-pci module, the user can create the VFs using the sysfs interface, and these VFs will be bound to vfio-pci module automatically.

When the PF is bound to vfio-pci, by default it will have a randomly generated VF token. For security reasons, this token is write only, so the user cannot read it from the kernel directly. To access the VFs, the user needs to create a new token, and use it to initialize both VF and PF devices. The tokens are in UUID format, so any UUID generation tool can be used to create a new token.

This VF token can be passed to DPDK by using EAL parameter --vfio-vf-token. The token will be used for all PF and VF ports within the application.

  1. Generate the VF token by uuid command

    14d63f20-8445-11ea-8900-1f9ce7d5650d

  2. Load the vfio-pci module with enable_sriov parameter set

    sudo modprobe vfio-pci enable_sriov=1

    Alternatively, pass the enable_sriov parameter through the sysfs if the module is already loaded or is built-in:

    echo 1 | sudo tee /sys/module/vfio_pci/parameters/enable_sriov

  3. Bind the PCI devices to vfio-pci driver

    ./usertools/dpdk-devbind.py -b vfio-pci 0000:86:00.0

  4. Create the desired number of VF devices

    echo 2 > /sys/bus/pci/devices/0000:86:00.0/sriov_numvfs

  5. Start the DPDK application that will manage the PF device

    <build_dir>/app/dpdk-testpmd -l 22-25 -n 4 -a 86:00.0 \
--vfio-vf-token=14d63f20-8445-11ea-8900-1f9ce7d5650d --file-prefix=pf -- -i

  6. Start the DPDK application that will manage the VF device

    <build_dir>/app/dpdk-testpmd -l 26-29 -n 4 -a 86:02.0 \
--vfio-vf-token=14d63f20-8445-11ea-8900-1f9ce7d5650d --file-prefix=vf0 -- -i

Note

Linux versions earlier than version 5.7 do not support the creation of virtual functions within the VFIO framework.

5.4.4. Troubleshooting VFIO

In certain situations, using dpdk-devbind.py script to bind a device to VFIO driver may fail. The first place to check is the kernel messages:

dmesg | tail
...
[ 1297.875090] vfio-pci: probe of 0000:31:00.0 failed with error -22
...

In most cases, the error -22 indicates that the VFIO subsystem could not be enabled because there is no IOMMU support.

To check whether the kernel has been booted with correct parameters, one can check the kernel command-line:

cat /proc/cmdline

Please refer to earlier sections on how to configure kernel parameters correctly for your system.

If the kernel is configured correctly, one also has to make sure that the BIOS configuration has virtualization features (such as Intel® VT-d). There is no standard way to check if the platform is configured correctly, so please check with your platform documentation to see if it has such features, and how to enable them.

In certain distributions, default kernel configuration is such that the no-IOMMU mode is disabled altogether at compile time. This can be checked in the boot configuration of your system:

cat /boot/config-$(uname -r) | grep NOIOMMU
# CONFIG_VFIO_NOIOMMU is not set

If CONFIG_VFIO_NOIOMMU is not enabled in the kernel configuration, VFIO driver will not support the no-IOMMU mode, and other alternatives (such as UIO drivers) will have to be used.

5.4.5. VFIO Platform

VFIO Platform is a kernel driver that extends capabilities of VFIO by adding support for platform devices that reside behind an IOMMU. Linux usually learns about platform devices directly from device tree during boot-up phase, unlike for example, PCI devices which have necessary information built-in.

To make use of VFIO platform, the vfio-platform module must be loaded first:

sudo modprobe vfio-platform

Note

By default vfio-platform assumes that platform device has dedicated reset driver. If such driver is missing or device does not require one, this option can be turned off by setting reset_required=0 module parameter.

Afterwards platform device needs to be bound to vfio-platform. This is standard procedure requiring two steps. First driver_override, which is available inside platform device directory, needs to be set to vfio-platform:

sudo echo vfio-platform > /sys/bus/platform/devices/DEV/driver_override

Next DEV device must be bound to vfio-platform driver:

sudo echo DEV > /sys/bus/platform/drivers/vfio-platform/bind

On application startup, DPDK platform bus driver scans /sys/bus/platform/devices searching for devices that have driver symbolic link pointing to vfio-platform driver. Finally, scanned devices are matched against available PMDs. Matching is successful if either PMD name or PMD alias matches kernel driver name or PMD name matches platform device name, all in that order.

VFIO Platform depends on ARM/ARM64 and is usually enabled on distributions running on these systems. Consult your distributions documentation to make sure that is the case.

5.4.6. Bifurcated Driver

PMDs which use the bifurcated driver co-exists with the device kernel driver. On such model the NIC is controlled by the kernel, while the data path is performed by the PMD directly on top of the device.

Such model has the following benefits:

  • It is secure and robust, as the memory management and isolation is done by the kernel.

  • It enables the user to use legacy linux tools such as ethtool or ifconfig while running DPDK application on the same network ports.

  • It enables the DPDK application to filter only part of the traffic, while the rest will be directed and handled by the kernel driver. The flow bifurcation is performed by the NIC hardware. As an example, using Flow isolated mode allows to choose strictly what is received in DPDK.

More about the bifurcated driver can be found in NVIDIA bifurcated PMD presentation.

5.4.7. UIO

Warning

Using UIO drivers is inherently unsafe due to this method lacking IOMMU protection, and can only be done by root user.

In situations where using VFIO is not an option, there are alternative drivers one can use. In many cases, the standard uio_pci_generic module included in the Linux kernel can be used as a substitute for VFIO. This module can be loaded using the command:

sudo modprobe uio_pci_generic

Note

uio_pci_generic module doesn’t support the creation of virtual functions.

As an alternative to the uio_pci_generic, there is the igb_uio module which can be found in the repository dpdk-kmods. It can be loaded as shown below:

sudo modprobe uio
sudo insmod igb_uio.ko

Note

For some devices which lack support for legacy interrupts, e.g. virtual function (VF) devices, the igb_uio module may be needed in place of uio_pci_generic.

Note

If UEFI secure boot is enabled, the Linux kernel may disallow the use of UIO on the system. Therefore, devices for use by DPDK should be bound to the vfio-pci kernel module rather than any UIO-based module. For more details see Binding and Unbinding Network Ports to/from the Kernel Modules below.

Note

If the devices used for DPDK are bound to a UIO-based kernel module, please make sure that the IOMMU is disabled or is in passthrough mode. One can add intel_iommu=off or amd_iommu=off or intel_iommu=on iommu=pt in GRUB command line on x86_64 systems, or add iommu.passthrough=1 on aarch64 systems.