>I seem to remember that AIX has a signal specifically for this
>situation (SIGDANGER?). I seem to remember that this came up
>last time the out of memory thread surfaced too.
Yep, that's what I meant. SIGDANGER is better, I think.
> The bottom line is that there is no single algorithm, simple
>or complex, that will consistently pick a non-stupid process
>to kill for everyone.
That's correct. This is not something that should go into the kernel,
because it's too complicated.
> Might I suggest that SIGDANGER be introduced and used to warn
>everything that they damn well better try and sort themselves
>out?
Yes.
This could even be a program that writes something to the system
console, so root can go in and kill processes. I'm dead against
setting ANY policy in the kernel for this.
Would it be possible to reserve some memory for root only, via
a tunable parameter, to allow at least "ps" and kill to work?
>If things don't get better we can try killing things - perhaps
>starting with things that don't have a SIGDANGER handler, followed
>by things that didn't reduce their footprint in response to
>SIGDANGER?
This borders on setting kernel policy, again.
-- Thomas Koenig, Thomas.Koenig@ciw.uni-karlsruhe.de, ig25@dkauni2.bitnet. The joy of engineering is to find a straight line on a double logarithmic diagram.