Re: [patch] Real-Time Preemption, -RT-2.6.10-rc2-mm2-V0.7.32-2

[john cooper <john.cooper@timesys.com>]

Gene Heskett wrote:

> ...I'd had that happen twice
> before and panic'd, even hitting the reset button once and doing a
> cold powerdown once, this before I noticed there was some disk
> activity going on, so I figured what the hell and walked away to go
> see what the missus was watching on the telly.  And when I came back
> half an hour later I was looking at a login prompt once I'd moved the 
> mouse.

Interesting to read the above.  I found similar
repeatable behavior when I launched a "make -j"
build on an arbitrary kernel tree when running
-V0.7.32-2.  I originally suspected the OOM
behavior to differ (somehow) with the 32-2 patch
compared to vanilla 2.6.10-rc2-mm3, but that was
a red herring.

I did notice the OOM killer nailing procs such as
sendmail, portmap, syslogd, lpd, etc..  fairly
early on in the build.  Eventually the login sh
got nailed and getty was respawned by init to play
again.  In oom_kill.c:badness(), 0 uid procs
should be given a limited stay, unless you run a
sprawling test load as root as I had.  This puts
everything into the same badness()-space and sys
daemons are equally likely to get wacked.

-john


-- 
john.cooper@timesys.com