An IOMMU may be present and enabled on many platforms. When an IOMMU is present and enabled, it is recommended that SPDK applications are deployed with the
vfio-pci kernel driver. SPDK's
scripts/setup.sh script will automatically select
vfio-pci in this case.
However, some devices do not function correctly when bound to
vfio-pci and instead must be attached to the
uio_pci_generic kernel driver. In that case, users should take care to disable the IOMMU or to set it into passthrough mode prior to running
To disable the IOMMU or place it into passthrough mode, add
intel_iommu=on iommu=pt to the GRUB command line on x86_64 system, or add
iommu.passthrough=1 on arm64 systems.
There are also some instances where a user may not want to use
uio_pci_generic or the kernel version they are using has a bug where
uio_pci_generic fails to bind to NVMe drives. In these cases, users can build the
igb_uio kernel module which can be found in dpdk-kmods repository. To ensure that the driver is properly bound, users should specify
One of the benefits of using the
VFIO Linux kernel driver is the ability to perform DMA operations with peripheral devices as unprivileged user. The permissions to access particular devices still need to be granted by the system administrator, but only on a one-time basis. Note that this functionality is supported with DPDK starting from version 18.11.
Make sure the target user has RW access to at least one hugepage mount. A good idea is to create a new mount specifically for SPDK:
Then start SPDK applications with an additional parameter
Full guide on configuring hugepage mounts is available in the Linux Hugetlbpage Documentation
VFIO device access is protected with sysfs file permissions and can be configured with chown/chmod.
Please note that the VFIO device isolation is based around IOMMU groups and it's only possible to change permissions of the entire group, which might possibly consist of more than one device. (You could also apply a custom kernel patch to further isolate those devices in the kernel, but it comes with potential risks as described on Alex Williamson's VFIO blog, with the patch in question available here: [PATCH] pci: Enable overrides for missing ACS capabilities)
Let's assume we want to use PCI device
0000:04:00.0. First of all, verify that it has an IOMMU group assigned:
The output should be e.g.
Which means that the device is a part of the IOMMU group 5. We can check if there are any other devices in that group.
In this case
0000:04:00.0 is an I/OAT channel which comes with 7 different channels associated with the same IOMMU group.
To give the user
spdk full access to the VFIO IOMMU group 5 and all its devices, use the following:
As soon as the first device is attached to SPDK, all of SPDK memory will be mapped to the IOMMU through the VFIO APIs. VFIO will try to mlock that memory and will likely exceed user ulimit on locked memory. Besides having various SPDK errors and failures, this would also pollute the syslog with the following entries:
The limit can be checked by running the following command as target user: (output in kilobytes)
On Ubuntu 18.04 this returns 16384 (16MB) by default, which is way below what SPDK needs.
The limit can be increased with one of the methods below. Keep in mind SPDK will try to map not only its reserved hugepages, but also all the memory that's shared by its vhost clients as described in the Vhost processing guide.
/etc/security/limits.conf file as root and append the following:
Then logout from the target user account. The changes will take effect after the next login.
Linux offers a
prlimit utility that can override limits of any given process. On Ubuntu, it is a part of the
Note that the above needs to be executed before the first device is attached to the SPDK application.