Name

The eCos synthetic target — Overview

Description

Usually eCos runs on either a custom piece of hardware, specially designed to meet the needs of a specific application, or on a development board of some sort that is available before the final hardware. Such boards have a number of things in common:

  1. Obviously there has to be at least one processor to do the work. Often this will be a 32-bit processor, but it can be smaller or larger. Processor speed will vary widely, depending on the expected needs of the application. However the exact processor being used tends not to matter very much for most of the development process: the use of languages such as C or C++ means that the compiler will handle those details.
  2. There needs to be memory for code and for data. A typical system will have two different types of memory. There will be some non-volatile memory such as flash, EPROM or masked ROM. There will also be some volatile memory such as DRAM or SRAM. Often the code for the final application will reside in the non-volatile memory and all of the RAM will be available for data. However updating non-volatile memory requires a non-trivial amount of effort, so for much of the development process it is more convenient to burn suitable firmware, for example RedBoot, into the non-volatile memory and then use that to load the application being debugged into RAM, alongside the application data and a small area reserved for use by the firmware.
  3. The platform must provide certain mimimal I/O facilities. Most eCos configurations require a clock signal of some sort. There must also be some way of outputting diagnostics to the user, often but not always via a serial port. Unless special debug hardware is being used, source level debugging will require bidirectional communication between a host machine and the target hardware, usually via a serial port or an ethernet device.
  4. All the above is not actually very useful yet because there is no way for the embedded device to interact with the rest of the world, except by generating diagnostics. Therefore an embedded device will have additional I/O hardware. This may be fairly standard hardware such as an ethernet or USB interface, or special hardware designed specifically for the intended application, or quite often some combination. Standard hardware such as ethernet or USB may be supported by eCos device drivers and protocol stacks, whereas the special hardware will be driven directly by application code.

Much of the above can be emulated on a typical PC running Linux. Instead of running the embedded application being developed on a target board of some sort, it can be run as a Linux process. The processor will be the PC's own processor, for example an x86, and the memory will be the process' address space. Some I/O facilities can be emulated directly through system calls. For example clock hardware can be emulated by setting up a SIGALRM signal, which will cause the process to be interrupted at regular intervals. This emulation of real hardware will not be particularly accurate, the number of cpu cycles available to the eCos application between clock ticks will vary widely depending on what else is running on the PC, but for much development work it will be good enough.

Other I/O facilities are provided through an I/O auxiliary process, ecosynth, that gets spawned by the eCos application during startup. When an eCos device driver wants to perform some I/O operation, for example send out an ethernet packet, it sends a request to the I/O auxiliary. That is an ordinary Linux application so it has ready access to all normal Linux I/O facilities. To emulate a device interrupt the I/O auxiliary can raise a SIGIO signal within the eCos application. The HAL's interrupt subsystem installs a signal handler for this, which will then invoke the standard eCos ISR/DSR mechanisms. The I/O auxiliary is based around Tcl scripting, making it easy to extend and customize. It should be possible to configure the synthetic target so that its I/O functionality is similar to what will be available on the final target hardware for the application being developed.

A key requirement for synthetic target code is that the embedded application must not be linked with any of the standard Linux libraries such as the GNU C library: that would lead to a confusing situation where both eCos and the Linux libraries attempted to provide functions such as printf. Instead the synthetic target support must be implemented directly on top of the Linux kernels' system call interface. For example, the kernel provides a system call for write operations. The actual function write is implemented in the system's C library, but all it does is move its arguments on to the stack or into certain registers and then execute a special trap instruction such as int 0x80. When this instruction is executed control transfers into the kernel, which will validate the arguments and perform the appropriate operation. Now, a synthetic target application cannot be linked with the system's C library. Instead it contains a function cyg_hal_sys_write which, like the C library's write function, pushes its arguments on to the stack and executes the trap instruction. The Linux kernel cannot tell the difference, so it will perform the I/O operation requested by the synthetic target. With appropriate knowledge of what system calls are available, this makes it possible to emulate the required I/O facilities. For example, spawning the ecosynth I/O auxiliary involves system calls cyg_hal_sys_fork and cyg_hal_sys_execve, and sending a request to the auxiliary uses cyg_hal_sys_write.

In many ways developing for the synthetic target is no different from developing for real embedded targets. eCos must be configured appropriately: selecting a suitable target such as i386linux will cause the configuration system to load the appropriate packages for this hardware; this includes an architectural HAL package and a platform-specific package; the architectural package contains generic code applicable to all Linux platforms, whereas the platform package is for specific Linux implementations such as the x86 version and contains any processor-specific code. Selecting this target will also bring in some device driver packages. Other aspects of the configuration such as which API's are supported are determined by the template, by adding and removing packages, and by fine-grained configuration.

In other ways developing for the synthetic target can be much easier than developing for a real embedded target. For example there is no need to worry about building and installing suitable firmware on the target hardware, and then downloading and debugging the actual application over a serial line or a similar connection. Instead an eCos application built for the synthetic target is mostly indistinguishable from an ordinary Linux program. It can be run simply by typing the name of the executable file at a shell prompt. Alternatively you can debug the application using whichever version of gdb is provided by your Linux distribution. There is no need to build or install special toolchains. Essentially using the synthetic target means that the various problems associated with real embedded hardware can be bypassed for much of the development process.

The eCos synthetic target provides emulation, not simulation. It is possible to run eCos in suitable architectural simulators but that involves a rather different approach to software development. For example, when running eCos on the psim PowerPC simulator you need appropriate cross-compilation tools that allow you to build PowerPC executables. These are then loaded into the simulator which interprets every instruction and attempts to simulate what would happen if the application were running on real hardware. This involves a lot of processing overhead, but depending on the functionality provided by the simulator it can give very accurate results. When developing for the synthetic target the executable is compiled for the PC's own processor and will be executed at full speed, with no need for a simulator or special tools. This will be much faster and somewhat simpler than using an architectural simulator, but no attempt is made to accurately match the behaviour of a real embedded target.