For a few days, I've been working on a new implementation of rtapi called "uspace". If you read the -commit message or lurk on the -devel IRC channel, you have seen some talk about this already.
"Uspace" uses POSIX threads, rather than the old libpth implementation. If the rtapi_app executable is setuid root, then it will enable I/O via iopl() and RT-PREEMPT style realtime via SCHED_FIFO. If it's not setuid root, then the functionality is roughly similar to what the old "sim" rtapi implementation supplied, running the realtime code but without giving any realtime guarantees or access to hardware. I have been testing on a Core 2 Duo desktop with Debian 7 and Debian's 3.2.0-rt-amd64 kernel. When I run latency-test with a 25us thread and a 1ms thread, my latencies are ~16us; when I run just a 1ms thread, my latencies are ~90us. (16us is about the same as cyclictest, not sure why I get a higher latency with just one thread though--this is something I'll be looking into) I haven't actually toggled any I/O bits yet, but in a case of famous last words, I figure that will "just work" once I build serport on it (the hardware has no parport). This work is distinct from Michael Buesch's work and from the work in machinekit. One specific bit of code is adapted pretty directly from ubc3, though -- a function called harden_rt(). I'm also glad that I had the ubc3 code to study as I wrote my implementation. The main "feature" of my implementation is that the branch will only contain this one topic, though I hope that it will be possible to base some other projects (hm2-eth and userspace pci come to mind) on top of my branch. I hope that by having a focused purpose, it will be uncontroversial to merge. By contrast, the work in machinekit branches such as the ubc3 branch is structured so that it is not possible to merge only their POSIX Threads implementation; a lot of other unrelated work would have to be merged at the same time, including work that some developers (including me) consider downright undesirable. My work introduces a C++ class interface to writing new userspace realtime layers. As written, the layer was able to cater to both pth and posix threads implementations (the Posix-specific stuff is around 150 lines out of about 950; 8 methods must be implemented for a new rtos), so it may be the case that this work would also ease porting to xenomai userspace or rtai lxrt; I have no personal interest in this work. Because it's not needed, the Pth implementation was deleted. Right now I reserve the right to rebase the branch frequently, so if you want to base work on that branch before it's marged to some branch with immutable history then please be familiar with how to manage your git workflow when upstream rebases (discussed in the git-rebase manpage). I hope that the rebasing will actually settle down once buildbot is happy with it, though. Thanks to those on IRC with kind words so far, and to cradek for some initial review of the code. (and catching why period couldn't be "unsigned long") Follow-ups to list please. Jeff ------------------------------------------------------------------------------ Open source business process management suite built on Java and Eclipse Turn processes into business applications with Bonita BPM Community Edition Quickly connect people, data, and systems into organized workflows Winner of BOSSIE, CODIE, OW2 and Gartner awards http://p.sf.net/sfu/Bonitasoft _______________________________________________ Emc-developers mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/emc-developers
