Porting — Adding support for other hosts


The initial development effort of the eCos synthetic target happened on x86 Linux machines. Porting to other platforms involves addressing a number of different issues. Some ports should be fairly straightforward, for example a port to Linux on a processor other than an x86. Porting to Unix or Unix-like operating systems other than Linux may be possible, but would involve more effort. Porting to a completely different operating system such as Windows would be very difficult. The text below complements the eCos Porting Guide.

Other Linux Platforms

Porting the synthetic target to a Linux platform that uses a processor other than x86 should be straightforward. The simplest approach is to copy the existing i386linux directory tree in the hal/synth hierarchy, then rename and edit the ten or so files in this package. Most of the changes should be pretty obvious, for example on a 64-bit processor some new data types will be needed in the basetype.h header file. It will also be necessary to update the toplevel ecos.db database with an entry for the new HAL package, and a new target entry will be needed.

Obviously a different processor will have different register sets and calling conventions, so the code for saving and restoring thread contexts and for implementing setjmp and longjmp will need to be updated. The exact way of performing Linux system calls will vary: on x86 linux this usually involves pushing some registers on the stack and then executing an int 0x080 trap instruction, but on a different processor the arguments might be passed in registers instead and certainly a different trap instruction will be used. The startup code is written in assembler, but needs to do little more than extract the process' argument and environment variables and then jump to the main linux_entry function provided by the architectural synthetic target HAL package.

The header file hal_io.h provided by the architectural HAL package provides various structure definitions, function prototypes, and macros related to system calls. These are correct for x86 linux, but there may be problems on other processors. For example a structure field that is currently defined as a 32-bit number may in fact may be a 64-bit number instead.

The synthetic target's memory map is defined in two files in the include/pkgconf subdirectory. For x86 the default memory map involves eight megabytes of read-only memory for the code at location 0x1000000 and another eight megabytes for data at 0x2000000. These address ranges may be reserved for other purposes on the new architecture, so may need changing. There may be some additional areas of memory allocated by the system for other purposes, for example the startup stack and any environment variables, but usually eCos applications can and should ignore those.

Other HAL functionality such as interrupt handling, diagnostics, and the system clock are provided by the architectural HAL package and should work on different processors with few if any changes. There may be some problems in the code that interacts with the I/O auxiliary because of lurking assumptions about endianness or the sizes of various data types.

When porting to other processors, a number of sources of information are likely to prove useful. Obviously the Linux kernel sources and header files constitute the ultimate authority on how things work at the system call level. The GNU C library sources may also prove very useful: for a normal Linux application it is the C library that provides the startup code and the system call interface.

Other Unix Platforms

Porting to a Unix or Unix-like operating system other than Linux would be somewhat more involved. The first requirement is toolchains: the GNU compilers, gcc and g++, must definitely be used; use of other GNU tools such as the linker may be needed as well, because eCos depends on functionality such as prioritizing C++ static constructors, and other linkers may not implement this or may implement it in a different and incompatible way. A closely related requirement is the use of ELF format for binary executables: if the operating system still uses an older format such as COFF then there are likely to be problems because they do not provide the flexibility required by eCos.

In the architectural HAL there should be very little code that is specific to Linux. Instead the code should work on any operating system that provides a reasonable implementation of the POSIX standard. There may be some problems with program startup, but those could be handled at the architectural level. Some changes may also be required to the exception handling code. However one file which will present a problem is hal_io.h, which contains various structure definitions and macros used with the system call interface. It is likely that many of these definitions will need changing, and it may well be appropriate to implement variant HAL packages for the different operating systems where this information can be separated out. Another possible problem is that the generic code assumes that system calls such as cyg_hal_sys_write are available. On an operating system other than Linux it is possible that some of these are not simple system calls, and instead wrapper functions will need to be implemented at the variant HAL level.

The generic I/O auxiliary code should be fairly portable to other Unix platforms. However some of the device drivers may contain code that is specific to Linux, for example the PF_PACKET socket address family and the ethertap virtual tunnelling interface. These may prove quite difficult to port.

The remaining porting task is to implement one or more platform HAL packages, one per processor type that is supported. This should involve much the same work as a port to another processor running Linux.

When using other Unix operating systems the kernel source code may not be available, which would make any porting effort more challenging. However there is still a good chance that the GNU C library will have been ported already, so its source code may contain much useful information.

Windows Platforms

Porting the current synthetic target code to some version of Windows or to another non-Unix platform is likely to prove very difficult. The first hurdle that needs to be crossed is the file format for binary executables: current Windows implementations do not use ELF, instead they use their own format PE which is a variant of the rather old and limited COFF format. It may well prove easier to first write an ELF loader for Windows executables, rather than try to get eCos to work within the constraints of PE. Of course that introduces new problems, for example existing source-level debuggers will still expect executables to be in PE format.

Under Linux a synthetic target application is not linked with the system's C library or any other standard system library. That would cause confusion, for example both eCos and the system's C library might try to define the printf function, and introduce complications such as working with shared libraries. For much the same reasons, a synthetic target application under Windows should not be linked with any Windows DLL's. If an ELF loader has been specially written then this may not be much of a problem.

The next big problem is the system call interface. Under Windows system calls are generally made via DLL's, and it is not clear that the underlying trap mechanism is well-documented or consistent between different releases of Windows.

The current code depends on the operating system providing an implementation of POSIX signal handling. This is used for I/O purposes, for example SIGALRM is used for the system clock, and for exceptions. It is not known what equivalent functionality is available under Windows.

Given the above problems a port of the synthetic target to Windows may or may not be technically feasible, but it would certainly require a very large amount of effort.