User-Space Device Drivers in Linux: A First Look
Mats Liljegren
Senior Software Architect

Device drivers in Linux are traditionally run in kernel space, but can
also be run in user space. This paper will take a look at running 
drivers in user space, trying to answer the questions in what degree
the driver can run in user space and what can be gained from this?

In the ‘90s, user-space drivers in Linux were much about how
to make graphics run faster[1] by avoiding calling the kernel.
These drivers where commonly used by the X-windows server.

User-space driver has become ever more important, as a blog
post by Tedd Hoff[2] illustrates. In his case the kernel is seen as
the problem when trying to achieve high server connection capacity.

Network interface hardware companies like Intel, Texas Instruments
and Freescale have picked up on this and are now providing
software solutions for user-space drivers supporting their hardware.

Device drivers normally run in kernel space, since handling
interrupts and mapping hardware resources require privileges
that only the kernel space is allowed to have. However, it is not
without drawbacks.

Each call to the kernel must perform a switch from user mode
to supervisor mode, and then back again. This takes time, which can
become a performance bottleneck if the calls are frequent.
Furthermore, the overhead is very much non-predictable, which
has a negative performance impact on real-time applications.

The kernel-space API is different. For example, malloc() needs
to be replaced by one of the several types of memory allocations
that the kernel can offer, such as kmalloc(), vmalloc(), alloc_pages()
or get_zeroed_page(). There is a lot to learn before becoming
productive.

The kernel-space API is less stable than user-space APIs, making
maintenance a challenge.

Debugging is very different in kernel space. Some tools often
used by user-space applications can be used for the kernel.
However, they represent exceptions rather than rule, where
LTTNG[3] is an example of the exception. To compensate for
this, the kernel has a lot of debug, tracing and profiling code
that can be enabled at compile time.

A crashing or misbehaving kernel tends to have a more severe
impact on the system than a crashing or misbehaving application,
which can affect robustness as well as how easy it is to debug.

The kernel space is a very different programming environment
than user space. It is more restricted, for example only C language
is supported. This rules out any script based prototyping.

If there are so many problems with having device drivers in
kernel space, is it time to have all drivers in user space instead?
As always, everything has its drawbacks, user-space drivers are
no exception.

Most of the issues with kernel-space drivers are solved by having
the driver in user space, but the issue with interface stability is
only true for very simple user-space drivers.

For more advanced user-space drivers, many of the interfaces
available for kernel-space drivers need to be re-implemented
for user-space drivers. This means that interface stability will still
be an issue.

There is a fear in the Linux kernel community that user-space
drivers are used as a tool to avoid the kernel‘s GPLv2 license.
This would undermine the idea with free open source software
ideas that GPLv2 has. However, this is outside the scope of
this paper.

Apart from this there are technical challenges for user-space
drivers.

Without question, interrupt handling is the biggest challenge
for a user-space driver. The function handling an interrupt is
called in privileged execution mode, often called supervisor
mode. User-space drivers have no permission to execute in
privileged execution mode, making it impossible for user-space
drivers to implement an interrupt handler.

There are two ways to deal with this problem: Either you do not
use interrupts, which means that you have to poll instead.
Or have a small kernel-space driver handling only the interrupt.
In the latter case you can inform the user-space driver of an
interrupt either by a blocking call, which unblocks when
an interrupt occurs, or using POSIX signal to preempt the
user-space driver.

Polling is beneficial if interrupts are frequent, since there is
considerable overhead associated with each interrupt, due to
the switch from user mode to supervisor mode and back that
it causes. Each poll attempt on the other hand is usually only
a check for a value on a specific memory address.

When interrupts become scarcer, polling will instead do a lot
of work just to determine that there was no work to do. This is
bad for power saving.

To get power saving when using user-space drivers with polling,
you can change the CPU clock frequency, or the number of
CPUs used, depending on work load. Both alternatives will
introduce ramp-up latency when there is a work load spike.

Many drivers use hardware dedicated to copying memory
areas managed by the CPU to or from memory areas managed
by hardware devices. Such dedicated hardware is called direct
memory access, or DMA. DMA relieves the CPU of such
memory copying.

There are some restrictions on the memory area used for
DMA. These restrictions are unique for each DMA device.
Common restrictions are that only a certain physical memory
range can be used, and that the physical memory range must be
consecutive.

Allocating memory that can be used for DMA transfers is
non-trivial for user-space drivers. However, since DMA
memory can be reused, you only need to allocate a pool of
memory to be used for DMA transfers at start-up. This means
that the kernel could help with providing such memory when
the user-space driver starts, but after that no further kernel
interactions would be needed.

Devices are often structured in a hierarchy. For example the
clock might be propagated in a tree-like fashion using different
dividers for different devices and offer the possibility to power
off the clock signal to save power.

There can be devices acting as a bridge, for example a PCI host
bridge. In this case you need to setup the bridge in order to
have access to any device connected on the other side of
the bridge.

In kernel space there are frameworks helping a device driver
programmer to solve these problems, but those frameworks
are not available in user space.

Since it is usually only the startup and shutdown phases that
affect other devices, the device interdependencies can be
solved by a kernel-space driver, while the user-space driver
can handle the actual operation of the device.

Network device drivers normally interfaces the kernel network
stack, just like block device drivers normally interfaces the kernel
file system framework.

User-space drivers have no direct access to such kernel services,
and must re-implement them.

The kernel has mechanisms for handling multiple clients
accessing the same resource, and for blocking threads waiting
for events or data from the device. These mechanisms are
available using standard interfaces like file descriptors, sockets,
or pipes.

To avoid using the kernel, the user-space driver needs to invent
its own interface.

The picture above shows how a user-space driver might be 
designed. The application interfaces the user-space part of the 
driver. The user-space part handles the hardware, but uses its 
kernel-space part for startup, shutdown, and receiving interrupts.

There are several frameworks and software solutions available 
to help designing a user-space driver.

There is a framework in the kernel called UIO [5][4] which facilitate 
writing a kernel-space part of the user-space driver. UIO has 
mechanisms for providing memory mapped I/O accessible for 
the user-space part of the driver.

The allocated memory regions are presented using a device 
file, typically called /dev/uioX, where X is a sequence number 
for the device. The user-space part will then open the file and 
perform mmap() on it. After that, the user-space part has direct 
access to its device.

By reading from the same file being opened for mmap(), the 
user-space part will block until an interrupt occurs. The content 
read will be the number of interrupts that has occurred. You can 
use select() on the opened file to wait for other events as well.

For user-space network drivers there are specialized solutions 
specific for certain hardware.

Data Plane Development Kit, DPDK[6], is a solution from Intel 
for user-space network drivers using Intel (x86) hardware. DPDK 
defines an execution environment which contains user-space 
network drivers. This execution environment defines a thread 
for each CPU, called lcore in DPDK. For maximum throughput 
you should not have any other thread running on that CPU.

While this package of libraries focuses on forwarding applications, 
you can implement server applications as well. For server DPDK 
applications you need to implement your own network stack 
and accept a DPDK specific interface for accessing the network.

Much effort has been put in memory handling, since this is often 
critical for reaching the best possible performance. There are 
special allocation and deallocation functions that try to minimize 
TLB[10] misses, use the most local memory for NUMA[11] 
systems and ensure even spread on multi-channel memory 
architectures [12]. 

User-space Data Plane Acceleration Architecture, USDPAA[7] , is 
a solution from Freescale for the same use case as DPDK but 
designed for their QorIQ architecture (PowerPC and ARM. 
The big difference is that QorIQ uses hardware for allocating, 
de-allocating and queuing network packet buffers. This makes 
memory management easier for the application.

TransportNetLib[8] is a solution from Texas Instruments. It 
is similar to USDPAA but for the Keystone architecture (ARM).

Open DataPlane, ODP[9], is a solution initiated by Linaro to do 
the same as DPDK, USDPAA and TransportNetLib, but with 
vendor generic interfaces.

To get the feeling for the potential performance gain from having 
a user mode network device driver, a DPDK benchmark application 
was designed and executed.

The design of the application can be seen in the picture above. 
It executes as four instances each running on its own CPU, or 
lcore, as DPDK calls them.

Each instance is dedicated to its own Ethernet device sending 
and receiving network packets. The packets sent has a magic 
word used for validating the packets and a timestamp used for 
measuring transport latency.

The instances are then paired using loopback cables. To be able 
to compare user-space driver with kernel-space driver, one 
pair accesses the hardware directly using the driver available in 
DPDK, and the other pair uses the pcap[13] interface. All four 
Ethernet devices are on the same PCI network card.

There is a fifth lcore (not shown in the picture above) which 
periodically collects statistics and displays it to the screen.

The hardware used was as follows: 
  o Supermicro A1SAi-2750F mother board using Intel Atom 
    C2750 CPU.  This CPU has 8 cores with no hyperthreading. 
  o 16GB of memory. 
  o Intel Ethernet server adapter i350-T4, 1000 Mbps.

The table below shows the throughput and latency for user-space 
driver compared to kernel-space driver.

A graph showing the throughput:

A graph showing the latency:

The theoretical throughput maximum is the sum of the send 
and receives speed for the network interface. In this case this 
is 1000 Mbps in each direction, giving a theoretical maximum 
of 2000 Mbps. The throughput includes packet headers and 
padding.

User-space driver achieved a throughput boost of about four 
times over kernel-space driver.

Latency was calculated by comparing the timestamp value found 
in the network packet with the current clock when packet was 
received. The latency for user-space driver was slightly less than 
for kernel-space driver.

Four threads, each continuously running netperf TCP streaming 
test against loop-back interface, were used as a stress while 
running the DPDK benchmark application. This had no noticeable 
impact on the measurements.

Implementing a user-space driver requires some work 
and knowledge. The major challenges are interrupts versus 
polling, power management and designing interface towards 
driver clients.

Support for user-space network drivers is a lot more developed 
than for other kinds of user-space drivers, especially for doing 
data plane forwarding type of applications.

A user-space driver can do everything a kernel-space driver 
can, except for implementing an interrupt handler.

Comparing a user-space network driver with a kernel-space 
network driver showed about four times better throughput 
for the user space driver. Latency did not show a significant 
difference.

The real-time characteristics should be good for user-space 
drivers since they do not invoke the kernel. This was not verified 
in this paper, though.