kicad-developers team mailing list archive

["Torsten Hüter" <torstenhtr@xxxxxx>]

Hi Wayne,

I've recently tried to compile KiCad on Windows (64Bit) and I've noticed that PcbNew is crashing at context switching - but I had no problems early 2015 with that. I've seen that you have already filed a bug report [1]. I did some research and this thing seems to cause constantly trouble, because some assembly "hackery" is required and only MASM is officially supported [2].

Rather than trying to fix Boost, makes it sense to consider an alternative way or library for the same purpose?

I've found so far these alternatives as back-end for "stackful" coroutines:

(1) Fibers on Windows

https://msdn.microsoft.com/de-de/library/windows/desktop/ms682661%28v=vs.85%29.aspx
https://en.wikipedia.org/wiki/Fiber_%28computer_science%29

This seems to be the safest alternative for Windows, because Fibers are fully supported by the OS and are thread-safe.

(2) POSIX ucontext.h - setcontext, getcontext, makecontext and swapcontext

Very similar compared to Windows Fibers, but deprecated since POSIX.1-2008, quote from https://en.wikipedia.org/wiki/Setcontext :

"[..] POSIX.1-2004 obsoleted these functions, and in POSIX.1-2008 they were removed [..]"

It is still possible to use these functions on Linux, but these are not future-proof and there are no replacement functions.

An interesting talk about this topic can be found here:
http://stackoverflow.com/questions/4298986/is-there-something-to-replace-the-ucontext-h-functions

(3) POSIX Threads

https://computing.llnl.gov/tutorials/pthreads/

It is possible to emulate Fibers with pthreads, here is a good article about this subject:

http://sjansen.blogspot.de/2008/04/coroutines-in-cc-beyond-duffs-device.html

The downside is the larger overhead, but this shouldn't be a big issue for KiCad, I'm sure that not more than a few coroutines are created for the tool framework. Quote from the above article:

"[..] As an interesting sidenote, I'm not sure how debuggers such as gdb and valgrind handle applications that switch their stacks. It might be that using threads actually makes things easier to debug in this case (as the tools already have support for threads). [..]"

Pthreads offer also some mechanisms for stack overflow detection.

(4) setjmp / longjmp

https://en.wikipedia.org/wiki/Setjmp.h

Context switching is also possible using these functions and some libraries like "picoro" use that for Coroutines. However stack handling is difficult, especially stack overflow protection is not easy to implement.

http://dotat.at/cgi/git/picoro.git

--

Summary: Fibers (Windows) and Pthreads (all POSIX-compliant operating systems) seem to be the most promising alternatives for context switching. This way it would be possible to get rid of the Boost.Context dependency while using standard libraries.

What do you think, should I invest more time into this topic (refactoring of coroutine.h using these alternatives and write a stress test application)?

Thanks, 
Torsten

[1] KiCad issues with Boost 1.59, bug report, https://github.com/Alexpux/MINGW-packages/issues/836
[2] Boost context requirements, http://www.boost.org/doc/libs/1_60_0/libs/context/doc/html/context/requirements.html