Jeremy would like a side-by-side comparison...
> and I wrote that 
> when running my stress-test on vanilla 2.6.6, the OOM-killer starts
> killing things off 'correctly' and it doesn't seem to crash 
> -- at least I haven't run it into the ground yet.  I suppose I can
> try again.
> 
> gds-2.6.7-r16 dies abruptly without ever calling the OOM-killer.
>  
> > If that's the case, would you be willing to test out 2.6.7-rc1 and
> > 2.6.7-rc2 and help narrow down when the lockup was introduced...
> 
> I assume you're speaking of the vanilla sources?
> I'll download and compile them....


Oh, yeah..  Fun stuff.

make -j all

start at 15:14
first OOM message at 15:15:48
first call trace at: 15:18:27

apache: page allocation failure. order:0, mode:0x1d2
Call trace:
	filemap_nopage
	do_no_page
	handle_mm_fault
	do_sparc64_fault
	sparc64_realfault_common

(I can send the full logs of this if you're really interested)

but the system kept trying to compile.  load average was about 273. heh
And sure, gcc reported lots of internal errors because of the OOM killer.

I could still swap consoles around, and then decided to killall -9 make &
cc1
the system fully recovered, and I was able to continue compiling with a bit 
nicer-on-the-system "make -j 5 all"

So from that respect, I couldn't make 2.6.6 crash.
Unlike gds-2.6.7-r16, which crashed hard.

vanilla 267-rc[12] continues compiling...