IndustryInsight

Real-Time Linux

Linux 2.6
for Embedded Systems—
Closing in on Real Time
While not yet ready for hard real-time computing, Linux 2.6 introduces
many new features that make it an excellent platform for a wide range
of embedded computing tasks

by Ravi Gupta, LynuxWorks

ith its low cost, abundant fea-

W tures and inherent openness,

Linux provides fertile ground
for creativity in embedded computing. As 1,000
Task Response Time Results
1GHz Pentium III operating under heavy interrupt load

its importance grows, we can even expect 24.18

100
Linux to become the platform where
progress first happens. The question is, 10
2.6
could Linux 2.6 be the breakthrough ver-
1
sion we’ve been anticipating for
embedded systems—the version that 0.1
opens the floodgates to Linux accep- 0.01
tance? The answer is “yes.” Best Average Worst
The embedded computing universe 24.18 kernel 0.073 1.133 156.66
is vast and encompasses computers of all 2.6 kernel 0.068 0.132 4.508
sizes, from tiny wristwatch cameras to
telecommunications switches with thou- Figure 1 Linux 2.6 shows marked improvement in response
sands of nodes distributed worldwide. time as loads are increased.
Embedded systems can be simple enough
to require only small microcontrollers, or
they may require massive parallel Taken together, these enhancements reliably. Linux is not yet a true real-time
processors with prodigious amounts of serve to both “firm up” Linux for operating system, yet Linux 2.6 intro-
memory and computing power. Linux 2.6 embedded computing while making it a duces improvements that make it a far
delivers enhancements to provide sup- more attractive alternative for a wider more worthy platform than in the past
port across the spectrum of these range of embedded applications. when responsiveness was an issue. The
embedded needs as well enhancements in three most significant improvements are:
the general areas of determinism, Firming Things Up
reduced memory contention, and the On the “firming up” front, timing • Preemption points in the kernel
leveraging of POSIX for thread creation constraints are endemic in many • An efficient scheduler
and management and task scheduling. embedded applications and must be met • Enhanced synchronization

Reprinted from November 2003

IndustryInsight

As with most general-purpose oper- moved to the expired queue. When this memory is “close” to that processor and
ating systems, Linux has always forbidden happens, the queues are switched: the other memory is “farther away.” This
a process scheduler from running when a expired queue becomes the current queue, means that when memory contention
process is executing in a call. Once a task and the empty current queue becomes the occurs, the nearer processor has superior
was in a system call, that task controlled the expired queue. As the tasks on the new rights to the memory. The 2.6 kernel pro-
processor until the call returned, regardless current queue have already been sorted, vides a set of functions that are aware of
of how long it took. This can cause more the scheduler can once again resume its the memory/processor topology. The
important tasks to be delayed while waiting simple algorithm of selecting the first task scheduler is able to use this information to
for the system call to complete. from the current queue. Not only is this favor tasks when they are using local
With Linux 2.6, the kernel is now procedure substantially faster than in the memory, thus ameliorating the memory
preemptible to a degree, making 2.6 more past, but it also works equally well contention bottleneck and improving
responsive than 2.4 and giving implemen- whether there are many tasks or just a few. throughput.
tors better control over the timing of A comparison of test response times
events. Moreover, in Linux 2.6, kernel between 2.4 and 2.6 is shown in Figure 1. Support For Custom Designs
code has been salted with preemption Until now, we have discussed funda-
points that enable the scheduler to run and Synchronization mental enhancements that make Linux 2.6
possibly block a current process so as to Oftentimes multiprocessing applica- better suited to embedded systems, par-
schedule a higher priority process. tions need to share resources including ticularly those systems requiring some
The 2.6 kernel can also be built with memory and devices. Programmers are degree of determinism and predictable
no virtual memory system, which solves a able to avoid race conditions by using a performance. In discussing how Linux
major sticking point in regard to respon- mutex function to ensure that only one task 2.6 supports large-scale multiprocessing
siveness. Software that has to meet dead- is using the resource at a time. Until now, systems and their applications, we have
lines is incompatible with virtual memory the Linux mutex implementation has also started to touch on how 2.6 can meet
demand paging, in which slow handling of always involved a system call to the kernel the needs of a wide spectrum of
page faults can ruin responsiveness. to decide whether to block the thread or embedded requirements in terms of
However, getting rid of virtual memory allow it to continue executing. But if the system size and configuration.
means the developer must ensure that decision happens to be to continue, why Pursuing this thread further, hard-
there will always be enough real memory make a time-consuming system call? ware designs for embedded systems are
available to meet application demands. The new implementation in Linux often customized for special applications
2.6 supports fast user-space mutexes, and it is common for designers to need to
Efficient Scheduling which check from user space to see if solve a design issue in an original way. For
Along with becoming somewhat pre- blocking is necessary and only perform example, a purpose-built board may use a
emptible, Linux 2.6 features a rewritten the system call when blocking the thread different IRQ management scheme than a
process scheduler developed to eliminate is required. These functions also use similar reference design. In order to run on
the slow algorithms of past versions. scheduling priority to decide which thread the new board, Linux has to be ported, or
Previously, the scheduler had to look at gets to execute when there is contention. altered to support the new hardware. This
each ready task and score its relative porting is made easier when the operating
importance. The time required for this Sharing Memory system is made up of components that are
algorithm varied with the number of tasks Large multiprocessor systems intro- well separated. Hence it is only necessary
and thus complex multitasking applica- duce proportionally large contention to change the code that actually needs to
tions would suffer from slow scheduling. issues. Embedded systems are sometimes change. The components of Linux 2.6 that
In Linux 2.6, the scheduler no longer extremely large systems of processors, are likely to be altered for a custom design
scans all tasks each time. Instead, when a as in telecom networks or mass storage have been refactored in line with a concept
task becomes ready to run it is sorted into systems. Here, numerous processors share called Subarchitecture.
position on a queue called the current memory, either as symmetric or loosely With Subarchitecture, components
queue. The scheduler only has to choose coupled multiprocessors. Symmetric mul- are clearly separate and can be individually
the task at the most favorable position on tiprocessing designs are inherently limited modified or replaced with minimal impact
the queue. Consequently, scheduling is because the shared memory has to be on other components of the board support
done in a constant amount of time. When equally accessible to all processors, package. This means more efficient devel-
the task is actually running, it is given a making competition for memory the lim- opment and faster time-to-implementation
period of time it may use the processor iting factor in both scalability and pro- and time-to-market for Linux-based
before it has to give way to another thread. cessing efficiency. embedded solutions (Figure 2).
When its time slice expires, the task is Linux 2.6 supports a different way of
moved to another queue, called the expired achieving multiprocessing, called Non Devices, Buses and I/O
queue, and sorted according to its priority. Uniform Memory Access (NUMA). With Linux is fast becoming the first
Eventually, all of the tasks on the NUMA, memory and processors are inter- choice of operating system for embedded
current queue will have been executed and connected, but for each processor some products aimed at consumer markets.

Reprinted from November 2003

 IndustryInsight

very large memory sizes have their choice

of 64-bit microprocessors with Linux 2.6.
The Importance of POSIX The Intel Itanium 64 architecture was
treated in a previous releases of Linux, and
There’s a lot to be said about the utility of POSIX. For example, the support continues in 2.6. Linux 2.6 also
POSIX standard describes a set of functions for thread creation and continues to cover the AMD64 architecture
management called POSIX threads, or pthreads. This functionality has with support of the AMD Opteron micro-
been available in past versions of Linux, but its implementation has processor. Nor is the PowerPC left out, as
been much improved in 2.6. The Native POSIX Thread Library (NPTL) PPC64 support is also available. Clearly, as
has been shown to be a significant improvement over the older
2.6 illustrates, the Linux community has
the momentum to keep up with innovations
LinuxThreads approach, and even improves other high-per formance
in large-bus, large-memory computing.
alternatives that have been available as patches.
Microcontrollers, on the other hand,
Along with POSIX threads, 2.6 provides POSIX signals and POSIX
have been something of a frontier for
high-resolution timers as part of the mainstream kernel. POSIX signals
Linux. Now they are supported on the
are an improvement over UNIX-style signals, which were the default in
mainstream Linux 2.6 kernel. In most
previous Linux releases. Unlike UNIX signals, POSIX signals cannot be cases, previous instances of Linux
lost and can carry information as an argument. Also, POSIX signals required a full-featured microprocessor
can be sent from one POSIX thread to another, rather than only from with a memory management unit (MMU).
process to process like UNIX signals. But simpler microcontrollers are typically
Embedded systems also often need to poll hardware or do other the more appropriate choice when low
tasks on a fixed schedule. POSIX timers make it easy to arrange any cost and simplicity are called for.
task to get scheduled periodically. The clock that the timer uses can There have been ways to put Linux on
be set to tick at a rate as fine as one kilohertz, so that software engi- MMU-less processors prior to version 2.6.
neers can control the scheduling of tasks with precision. The Linux for Microcontrollers project has
been a successful branch of Linux for some
important small systems. Version 2.6 inte-
grates a significant portion of uClinux into
This is largely due to its cost-effectiveness user interface and sometimes with no the production kernel, bringing microcon-
in what are often low-margin commodity operator interface. While previous Linux troller support into the Linux mainstream.
devices. Linux 2.6 delivers support for versions made it possible to build a head- The Linux 2.6 version supports sev-
several technologies that are key to the less system, some of the support software eral current microcontrollers that don’t
success of many types of consumer prod- was not removable, giving the kernel more have memory management units. These
ucts. For example, 2.6 includes the bulk than was necessary or desirable. include Motorola m68k processors such as
Advanced Linux Sound Architecture, or Linux 2.6 however can be configured to Dragonball and ColdFire, as well as
ALSA. This state-of-the-art facility sup- entirely omit support for unneeded dis- Hitachi H8/300 and NEC v850 micro-
ports USB and MIDI devices with fully plays, keyboards or mice. processors. The ETRAX family of net-
thread and multiprocessor-safe software. For portable products, Linux 2.6 working microcontrollers by Axis
With ALSA, a system can run multiple debuts the Bluetooth wireless interface, Communications is also supported.
sound cards, play and record at the same which is now taking its place next to As a caveat, Linux running on MMU-
time or mix multiple audio streams. 802.11 as a protocol option for wireless less processors will still be multitasking,
USB 2.0 also makes its debut on Linux communications. With both the SCO but will obviously not have the memory
2.6. We can expect that high-speed devices datalink for audio and the L2CAP for con- protection provided on fully endowed
will proliferate in the near future, and that nection-oriented data transfers available, processors. Consistent with the lack of true
Linux will be a leading platform for USB Linux 2.6 is an excellent choice wherever processes on these small platforms, there is
2.0 products. no-fuss, short-range wireless connectivity also little in the way of security.
Video4Linux, the system for sup- is a key requirement. In the area of middleweight systems,
porting video, is also new in Linux 2.6. sometimes an embedded system may use a
Although it is not backward compatible Size Matters conventional Intel-architecture micro-
with the previous video paradigm, it is On the other end of the spectrum are processor, but may need to manage more
intended for the latest generation of radio computers that provide exceptionally large RAM that can be addressed in the usual 32-
and TV tuners, video cameras and other resources, such as very large memory sizes bit address space. Intel’s Physical Address
multimedia devices. or high-throughput multiprocessors. These Extension (PAE) makes it possible for 32-
heavyweights have numerous embedded bit computers to access up to 64 Gbytes of
Headless Operation applications, such as mass data storage sys- memory through page frames. With PAE,
Deeply embedded systems have their tems and special compute engines. Linux 2.6 systems running on newer x86
own challenge and are often built with no Embedded Linux developers who need machines view memory through a movable

Reprinted from November 2003

IndustryInsight

“window,” allowing the system to Linux 2.6 Developers also need to

address up to 64 Gbytes of RAM. High-Level Components bear in mind that while Linux
Linux 2.6 support of PAE is espe- 2.6 is not a true real-time oper-
cially useful on systems that per- User Space ating system, it’s possible to see
form store-and-forward service App1 App2 ... one from there. That is, should
for images, video and applica- the need arise, the time and
tions where large datasets must be effort spent developing an
handled quickly. Kernel Network Device USB2.0/1394/ALSA ... embedded application on Linux
Scheduler 0(1) VM FS Thread Lib Others ... can be leveraged into the real-
Looking Ahead time space with the choice of a
Linux is easily the fastest Linux-compatible, POSIX-
Architecture/Platform (32/64 Bits)
growing operating system in the compliant RTOS. Hence
embedded world and Linux 2.6 x86 PPC ARM MIPS Others embedded developers can turn
will undoubtedly accelerate that to Linux 2.6 with the confi-
growth. Still critical to success of Hardware dence that they can go real time,
individual embedded projects, in almost no time.
however, is expertise in Figure 2 The increased modularity of the Linux 2.6 kernel
embedded design and makes it easier to create custom configurations for LynuxWorks
Linux/POSIX (see sidebar: “The specific embedded applications. San Jose, CA.
Importance of POSIX”, p. 33). (408)979-3900.
There is a lot on the way of [www.lynuxworks.com].
expert support and guidance that needs to it is an opportunity for competitive differ-
be wrapped around an embedded Linux entiation and time-to-market advantage
implementation. But with the proper for the embedded project.
vendor, this is not so much a challenge as

Reprinted from November 2003