Re: Keyboard lockup with NetBSD5

["Michael L. Hitch" <mhitch%lightning.msu.montana.edu@localhost>]

On Wed, 11 Nov 2009, Frank Wille wrote:

> John Klos wrote:
>
> > Was this inadvertent fallout from the timer updates?
>
> Which timer updates?
> You mean it was caused by the new timecounter code, we fixed a few weeks
> ago? No.

  Actually, I think the timecounter code could have affected this.  The
delay() routine was completely rewritten.  I'm not sure how accurate the 
old version was, or how accurate the new one is.

  Re that 2000 usec delay used apparently came from me when the Amiga 1200 
support was added.  Presumably that was needed to make the 1200 I had 
access to at the time, and hasn't caused any problems until NetBSD 5.0,
which had the new timecounter code.

  Another new thing in NetBSD 5.0 which may have a affect is that the 
handling of spl*() and IPL levels was revised.  I've not been able to use 
a serial port to connect to another machine's serial console [although I
have verified that the Amiga serial console does still work OK].  One 
thing in particular I think may be causing problems is the mutex 
implementation.  The mutex_enter_spin() will raise the IPL level (or leave 
it alone if the requested IPL is lower than the current level), but the 
mutex_exit() won't lower the IPL until the outermost spin mutex is 
released.  This means that if something enters a mutex with IPL_VM, 
followed by a short code segment that enters and exits a mutex with 
IPL_SCHED or IPL_HIGH, the IPL will remain at IPL_SCHED or IPL_HIGH until 
the mutex_exit is called that entered with IPL_VM.  This means code can 
possibly be running at a higher IPL longer than expected.  I'm not sure 
this could affect the kbd interrupt handling or not.

--
Michael L. Hitch                        mhitch%montana.edu@localhost
Computer Consultant
Information Technology Center
Montana State University        Bozeman, MT     USA