On Sun, Jun 29, 2014 at 12:29:50PM +0200, Marius Liebenberg wrote: > Hi Jeff, > I have been following your thread with great interest. I am not to clued > up with the different flavours of kernels and RTAI interfaces though. > Would it be possible to give us laymen a simple description of what your > work means to us, what it intends to achieve and what we have to look > for when testing it. I mean a real high level description please.
Short version: These patches enable linuxcnc to run with a different realtime kernel, RT-PREEMPT. A compatible kernel is packaged by Debian, which means less work by linuxcnc developers to run on a new OS version. Because it has higher latency than RTAI on most systems, this kernel isn't good for software step generation, but is good for use with smart FPGA boards. Since this version runs realtime code in userspace, it may make a good basis for drivers that work with ethernet motion control cards like the Mesa 7i80 family. The RT-PREEMPT kernel's home page is https://rt.wiki.kernel.org/index.php/Main_Page Long version: There are multiple choices for realtime kernels on Linux, and they each have trade-offs. For a very long time, the linuxcnc project has only been able to work with the "RTAI" realtime kernels. On a wide range of hardware (but by no means all hardware), "RTAI" provides realtime performance that is suitable for software step generation and in a few cases software encoder counting. However, at times the RTAI project has lagged far behind in kernel version support as well as availability on non-i386 architectures. Also, no linux distro offers a RTAI kernel, so the linuxcnc project has to build one themselves. This requirement, more than any other dependency of linuxcnc, is the reason that we are slow to offer packages for new operating system versions. Another realtime kernel is the RT-PREEMPT kernel. This project seems to have a lot more traction (though it also lags behind the mainline of kernel development), and more importantly, Debian offers "-rt" kernels with RT-PREEMPT patches. (Ubuntu seems to also offer -rt and/or -realtime kernels, but I am less familiar with the details on Ubuntu) Generally, the realtime performance of RT-PREEMPT is too slow for software step generation, so it's likely only to be useful for systems where you have motion control hardware like the various FPGA cards that work with linuxcnc. We couldn't switch to *only* RT-PREEMPT without leaving some of our users behind. My project is about supporting the RT-PREEMPT kernel (in addition to the RTAI kernel). This has all happened before, most recently by the people now working on the "machinekit" project, a linuxcnc derivative. The main problem with adopting their work into linuxcnc was that work that many linuxcnc developers wanted (support for RT-PREEMPT kernels) was mixed with work that at least a few developers (including me) did not want (such as an unfinished web interface to linuxcnc). Our inability to resolve this impasse is in part why the machinekit people are now choosing to go their own way. My work builds on the existing userspace simulator, sim_rtapi, but changes it to use the standard Unix API for multithreaded software, pthreads. It also provides rtapi versions of the kernel functions used in the hal components "serport", "hostmot2", "hm2_pci" and "hm2_test", so that these hardware drivers can now be loaded either in traditional RTAI kernel realtime systems and in the new RT-PREEMPT userspace realtime system. The framework should allow other hardware drivers to be adapted too, though for smart parallel (EPP) boards some additional groundwork will be required. So far, I've only adapted drivers for hardware I own and can personally test. Before talking a little bit more about what I find unfavorable in the machinekit code, I want to again emphasize that their code was very useful to study as I worked on this project, and that two specific parts (harden_rt and rtapi_pci) are adaptations of their GPL-licensed work, rather than my own original work. First off, as I touched on earlier, the branch that they proposed to merge into linuxcnc.org's master branch (unified-build-candidate-3) included a number of unrelated projects. The organization of their branch meant that we could choose either to merge all of it, or to merge none of it. Ultimately, this meant that no developers could give a resounding "yes" to the merge, and some developers (including me) had to come down on the side of "no" due to the unwanted parts. Git gives this summary of the changes in unified-build-candidate-3: 1412 commits, 645 files changed, 89622 insertions(+), 10373 deletions(-) while my branch, which has a focused purpose (what git calls a "topic branch"): 29 commits, 93 files changed, 4146 insertions(+), 2251 deletions(-) and it is organized in three separate phases which could be merged in three steps if that's how review proceeds. Second, there's a specific kind of bad design to the kernel-and- userspace portability layer in machinekit: they ended up adding a lot more #ifs to the driver code. One example, from the motenc driver (which admittedly I have not yet converted): diff --git a/src/hal/drivers/hal_motenc.c b/src/hal/drivers/hal_motenc.c --- a/src/hal/drivers/hal_motenc.c +++ b/src/hal/drivers/hal_motenc.c @@ -291,8 +301,11 @@ rtapi_app_main(void) i = 0; // Find a MOTENC card. - while((i < MAX_DEVICES) && ((pDev = pci_get_device(MOTENC_VENDOR_ID, MOTENC_DEVICE_ID, pDev)) != NULL)){ - +#if defined(BUILD_SYS_USER_DSO) + while((i < MAX_DEVICES) && ((pDev = rtapi_pci_get_device(MOTENC_VENDOR_ID, MOTENC_DEVICE_ID, pDev)) != NULL)){ +#else + while((i < MAX_DEVICES) && ((pDev = pci_get_device(MOTENC_VENDOR_ID, MOTENC_DEVICE_ID, pDev)) != NULL)){ +#endif // Allocate memory for device object. pDevice = hal_malloc(sizeof(Device)); There are many such #if defined checks, each one turning what should have been one line into 5. In my design, the function rtapi_pci_get_device exists whether you are in kernel or user, so hardware drivers can simply call the one required function. Some conditional compilation is required in a header, but the use sites all look nice and clean. In my branch, assuming I implemented rtapi_pci_get_device yet, the change would be - while((i < MAX_DEVICES) && ((pDev = pci_get_device(MOTENC_VENDOR_ID, MOTENC_DEVICE_ID, pDev)) != NULL)){ + while((i < MAX_DEVICES) && ((pDev = rtapi_pci_get_device(MOTENC_VENDOR_ID, MOTENC_DEVICE_ID, pDev)) != NULL)){ i.e., just prefix with "rtapi_". 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 Emc-developers@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/emc-developers
