On Monday, July 14, 2003, at 02:50  PM, Richard Earnshaw wrote:

> In trying to port the generic soft-ints code to the Integrator board 
> I've
> been running into a race condition that can lead to particular 
> interrupts
> becoming permanently blocked, even though the SPL level is 0 and the
> processor is sitting in the idle loop.  The main symptom is that
> XXX_ipending has a bit set, and the corresponding bit in intr_enabled 
> is
> zero, so the interrupt is permanently blocked.  Here is how I think 
> this
> can happen:

Yah, your analysis is right... However, if you take a look at the 
XScale ports that use it, you'll see that they already take care of the 
problem.  See line 515 of i80321_icu.c for an example of a platform 
that tracks softintrs using software only, or line 504 of becc_icu.c 
that uses hardware assist for softintrs.

Patching it as you suggest could also work (and, depending on the 
hard/soft interrupt glue, could be more efficient).

         -- Jason R. Thorpe <thorpej@wasabisystems.com>