Deadlocks with recent SMP current

[Scott Long]

Lukas Ertl wrote:
> On Fri, 13 Aug 2004, Robert Watson wrote:
> 
>> will eventually get a solid hang.  I tried it on a new SMP box with an 
>> NMI
>> button I received yesterday but was unable to get into the debugger.  I'm
>> in the process of de-obfuscating the NMI path to increase the chances of
>> successfully getting into the debugger and then I'll try again to see 
>> what
>> I can figure out.
> 
> 
> I have an NMI-enabled SMP box too, and the only message I got when 
> sending an NMI to the deadlocked system was 'kernel trap 12 with 
> interrupts disabled', but nothing more.
> 
> If you want me to test something, feel free to shout out.
> 
> cheers,
> le
> 

The NMI handler(s) try to do far too much and often get caught up in
mutexes.  I've locally changed it to just call either kdb_backtrace() or 
kdb_enter() and then reboot.  We might want to consider looking for
a general solution for this in 5.3.

Scott