On Fri, Nov 20, 2009 at 03:36:36AM +0000, Mindaugas Rasiukevicius wrote: > In fact, I am not convinced that such IPL tracking is very useful - we have > reduced amount of IPLs, which is much simpler model than we used to have, and > the problem you were recently debugging was miss-interpretation of certain > spin-mutex behaviour.
You have not described the bug itself, or the cause of the bug, but the cause of the difficulty in finding the cause of the bug. :-) And you're doing so with the benefit of hindsight. The bug was a panic, "cpu_switchto: switching above IPL_SCHED (8)", that typically occurred in the tty code. The reason was not clear until evidence provided by SPLDEBUG pointed to a shortcut/optimization of spin mutexes that made my new code interact with tty locking in unexpected ways. After painstakingly debugging a program, I find that it is easy to say, "The answer is obvious, now. It should have been obvious all along. It will be obvious the next time, for sure." I am tempted to delete all of the instrumentation. But then I am just as ill-prepared for the next "obvious" bug, and so are the developers who I work with. SPLDEBUG is single-purpose. It also has some bugs, which I am happy to describe. But let's keep it until we come up with something better. Dave -- David Young OJC Technologies dyo...@ojctech.com Urbana, IL * (217) 278-3933
