Friday, May 18, 2012

Technical Debt and RTEMS

Dr. Dobb's recently had an interview with Ward Cunningham who developed the first wiki among other notable contributions. The interview is interesting and I recommend reading it. But I wanted to pass along some thought on one term that resonated with me.
Technical Debt: Cunningham uses this term to refer to work that needs to be done before a desired change can be implemented or work required to propagate a desired change across a codebase. In the RTEMS world, we have multiple examples of this. 
The most common case of RTEMS technical debt is when a single change must be implemented across all or a set of BSPs at the same time. Recently, Jennifer and I converted the MIPS port from Simple Vectored (SV) to Programmable Interrupt Controller (PIC) interrupt model. We did this because the MIPS/Malta we were developing a BSP for had a more complicated interrupt structure than previous MIPS boards. It was logical to use the PIC rather than the SV model on this BSP. But to ensure that all MIPS BSPs were consistent, we had to implement the same change for six other BSPs.

There are cases with RTEMS where some preparatory work must be done before something else is implemented. A prominent case of this was the CPU Scheduler Plugin Framework work by Gedare Bloom. Before the plugin framework existed, the RTEMS CPU Scheduler was bits and pieces of code embedded in the places where threads changed states and priorities. The plugin framework captured those decision points and made it possible to change the scheduling algorithm at application time configuration (e.g. confdefs.h in RTEMS terms). This was refactoring and clean up needed to be able to implement SMP support for RTEMS.

There are also cases where both types of debt must be paid within a single area. The RTEMS file system infrastructure has evolved over the past few years. Sometimes, there are desirable changes which must be propagated across all file system implementations and for other changes which have required cleaning up an area before refactoring or reworking it.

Not paying your technical debt can lead to long-term pain on a project. The most prominent case of this with RTEMS is when the PowerPC was converted some SV to PIC interrupt model. In contrast to what Jennifer and I did for the MIPS, this time only some BSPs were converted to the PIC model. This meant there were two different BSP interrupt models that co-existed for years. Worse, they were not named at the time and were known as "old" and "new" for years. And if that wasn't bad enough, the method and type names were not the same as the implementation for the x86. It has taken years of paying technical debt to clean up this mess.

It is also something that RTEMS developers would like to avoid repeating. So when you submit a patch and you are asked to modify some code you don't care about to make it consistent or told to be consistent with another BSP, remember that we are just asking you to avoid incurring technical debt. After all, it will probably be someone else who has to pay it. Better to avoid it altogether.

No comments:

Post a Comment