Severe deadlock issues with 5.0/MP

[Anders Lindgren <ali%df.lth.se@localhost>]

I have a 6-way E3000 I was hoping to run NetBSD on with the recent SMP 
successes, so I recently installed a 5.0_BETA built on Jan 24, just 
shortly before the RC1 tag hit the tree it seems.

The system is running off of a 2x33G RAID1 Raidframe set, ffs logging is 
not yet enabled.

In order to test the MP robustness, I started building the latest RC1 
sources like so:

./build.sh -j 8 [ -M ... ] tools kernel=GENERIC.MP distribution > 
/tmp/mk.log 2>&1 &

..and sat back to watch the progress with tail -f. Before building gcc 
finished (still at building "tools"), all output just stopped. Pressing ^C 
on the "tail" just yielded "^C" in the terminal window (ssh session) but 
no effect, and after that the terminal session was completely dead.

To see whether this was a general issue or some race caused only by 
concurrent loads, I hooked up a serial console (so I could catch the 
output of a panic, if any, and maybe drop into ddb) and restarted the 
build without -j. The result was the same on the serial console however; 
after a few minutes of building, all output stops. Pressing return a few 
times did scroll the console contents, but when ^C:ing the "tail", I only 
get a ^C on the console, no prompt, and beyond that point the box is stone 
dead. Doesn't even respond to serial console break. Looks almost like a 
complete deadlock with interrupts blocked or something. No panic occurs.

Is anyone else having success with MP-kernels, or seeing similar problems?

I'll try a DIAGNOSTICS/LOCKDEBUG kernel and see what gives -- is there any 
other way I can get useful debugging out of it? (e.g. developer access to 
the serial console could be arranged; I just won't have on-the-fly access 
to power-cycle it if it hangs, due to its location)

Best regards,
ali:)