ixy is a simple userspace packet processing framework. It takes exclusive control of a network adapter and implements the whole driver in userspace. Its architecture is similar to DPDK and Snabb and completely different from (seemingly similar) frameworks such as netmap, pfq, pf_ring, or XDP (all of which rely on kernel components). In fact, reading both DPDK and Snabb drivers was crucial to understand some parts of the Intel 82599 datasheet better.

Check out the draft of our paper or watch the recording of our talk at 34C3 to learn more.

ixy is designed for educational purposes to learn how a network card works at the driver level. Lack of kernel code and external libraries allows you to look through the whole code from startup to the lowest level of the driver. Low-level functions like handling DMA descriptors are rarely more than a single function call away from your application logic.

A whole ixy app, including the whole driver, is only ~1000 lines of C code. Check out the ixy-fwd and ixy-pktgen example apps and look through the code. The code often references sections in the Intel 82599 datasheet or the VirtIO specification, so keep them open while reading the code. You will be surprised how simple a full driver for a network card can be.

Don't like C? We also have implementations in other languages including Rust, Go, Haskell, and OCaml.

• Driver for Intel NICs in the ixgbe family, i.e., the 82599ES family (aka Intel X520)
• Driver for paravirtualized virtio NICs
• Less than 1000 lines of C code for a packet forwarder including the whole driver
• No kernel modules needed (except vfio-pci when using the IOMMU / VFIO)
• Can run without root privileges (when using the IOMMU / VFIO)
• IOMMU support (see Using the IOMMU / VFIO)
• Simple API with memory management, similar to DPDK, easier to use than APIs based on a ring interface (e.g., netmap)
• Support for multiple device queues and multiple threads
• Super fast, can forward > 25 million packets per second on a single 3.0 GHz CPU core
• Super simple to use: no dependencies, no annoying drivers to load, bind, or manage - see step-by-step tutorial below
• BSD license

Tested on an Intel 82599ES (aka Intel X520), X540, and X550. Might not work on all variants of these NICs because our link setup code is a little bit dodgy.

Check out the draft of our paper.

If you prefer to dive into the code: Start by reading the apps in src/app then follow the function calls into the driver. The comments in the code refer to the Intel 82599 datasheet (Revision 3.3, March 2016).

Your NIC has full DMA access to your memory. A misconfigured NIC will cause memory corruptions that might crash your server or even destroy your filesystem. Do not run this on any systems that have anything remotely important on them if you want to modify the driver. Our version is also not necessarily safe and might be buggy. You have been warned.

Running ixy will unbind the driver of the given PCIe device without checking if it is in use. This means the NIC will disappear from the system. Do not run this on NICs that you need. We currently have a simple check if the device is actually a NIC, but trying to use another device could crash your system.

1. Install the following dependencies

   • gcc >= 4.8
   • make
   • cmake

   Run this on Debian/Ubuntu to install them:

   sudo apt-get install -y build-essential cmake
   

2. Configure 2MB hugepages in /mnt/huge using our script:

   cd ixy
   sudo ./setup-hugetlbfs.sh
   

3. Run cmake and make

   cmake .
   make
   

4. That's it! You can now run the included examples, e.g.:

   sudo ./ixy-pktgen 0000:XX:YY.Z
   

   Replace the PCI address as needed. All examples expect fully qualified PCIe bus addresses, i.e., typically prefixed with 0000:, as arguments. You can use lspci from the pciutils (Debian/Ubuntu) package to find the bus address. For example, lspci shows my 82599ES NIC as

   03:00.0 Ethernet controller: Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)

   which means that I have to pass 0000:03:00.0 as parameter to use it.

The usage of the IOMMU via the vfio-pci driver is implemented for ixgbe devices (Intel X520, X540, and X550). To use it, you have to:

0. Enable the IOMMU in the BIOS. On most Intel machines, the BIOS entry is called VT-d and has to be enabled in addition to any other virtualization technique.

1. Enable the IOMMU in the linux kernel. Add intel_iommu=on to your cmdline (if you are running a grub, the file /etc/default/grub.cfg contains a GRUB_CMDLINE_LINUX where you can add it).

2. Get the PCI address, vendor and device ID: lspci -nn | grep Ether returns something like 05:00.0 Ethernet controller [0200]: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 [8086:1528] (rev 01). In this case, 0000:05:00.0 is our PCI Address, and 8086 and 1528 are the vendor and device id, respectively.

3. Unbind the device from the ixgbe driver. echo $PCI_ADDRESS > /sys/bus/pci/devices/$PCI_ADDRESS/driver/unbind

4. Enable the vfio-pci driver. modprobe vfio-pci

5. Bind the device to the vfio-pci driver. echo $VENDOR_ID $DEVICE_ID > /sys/bus/pci/drivers/vfio-pci/new_id

6. Chown the device to the user. chown $USER:$GROUP /dev/vfio/*

7. That's it! Now you can compile and run ixy as stated above!

It's not the plan to implement every single feature, but a few more things would be nice to have. The list is in no particular order.

NICs that rely too much on firmware (e.g., Intel XL710) are not fun, because you end up only talking to a firmware that does everything. The same is true for NICs like the ones by Mellanox that keep a lot of magic in kernel modules, even when being used by frameworks like DPDK.

Interesting candidates would be NICs from the Intel igb and e1000e families as they quite common and reasonably cheap.

PCIe devices are attached to a specific CPU in NUMA systems. DMA memory should be pinned to the correct NUMA node. Threads handling packet reception should also be pinned to the same NUMA node.

NUMA handling must currently be done via numactl outside of ixy. Implementing it within ixy is annoying without depending on libnuma, so it's not implemented here.

What's the point of having multiple rx queues if there is no good way to distribute the traffic to them? Shouldn't be too hard. See Section 7.1.2.8 in the datasheet.

A simple rx-only app that writes packets to a .pcap file based on mmap and fallocate. Most of the code can be re-used from libmoon's pcap.lua.

A mempool can currently only be accessed by a single thread, i.e., a packet must be allocated and free'd by the same thread. This prevents apps that implement different processing steps on different cores, a common multi-threading model for complex chains.

The current limitation is the list of free buffers in the mempool, it's a simple stack at the moment. Replacing this with a lock-free stack or queue makes it safe for use with multiple threads.

The good news is that multi-threaded mempools are essentially the same problem as passing packets between threads: both can be implemented with the same queue/ring buffer data structure.

It's the lowest common denominator that everyone should be able to understand -- this is for educational purposes only. I've taken care to keep the code simple and understandable.

We are working on implementations in different languages including Rust, Go, Haskell, and OCaml.

We are currently facing a weird problem that impacts performance if your CPU is too fast. DPDK had the same problem in the past. Try applying bidirectional traffic to the forwarder and/or underclock your CPU to speed up ixy.

ixgbe_type.h is copied from the Intel driver, it's only used as a machine-readable version of the datasheet. ixy only uses #define definitions for registers and the two relatively simple structs for the DMA descriptors. Overall, ixy uses less than 100 lines of the file and we could remove the remainder.

But it's nice to have all the struct definitions right there when implementing a new driver feature. Copy & pasting magic values from the datasheet is significantly less fun. Another interesting approach to making these values available is writing a parser for the tables in the datasheet. Snabb does this.

No.

To understand how a NIC works and to understand how a framework like DPDK or Snabb works.