kernel killing processes when out of swap

[David Schultz]

On Tue, Apr 12, 2005, Steven Hartland wrote:
> ----- Original Message ----- 
> >There is no "large process detection".  The first process that tries
> >to fault in a new page after the system runs out of swap gets killed.
> 
> That makes sense. Me trying to connect to see what was going
> on would hence cause sshd to be killed ( not good ).
> It would seem that we could do with doing something to prevent this.
> We have had a few suggestions so far.
> 1. Using madvise, requires code changes per app

Using madvise doesn't require changes per app, since MADV_PROTECT
is inherited across exec.  You just have to write a wrapper, much
in the spirit of nice(1), to execute a protected version of X.

I agree that it's possible to do better, but this has already been
discussed several times before, and whenever it comes up, there's
always lots of noise because lots of people who don't understand
deadlock chime in.  If someone really wants to fix this in a
reliable way, they need to go off and implement the necessary
accounting in the VM system so that we don't overcommit swap.
Most of the other suggestions, such as SIGDANGER, are just heuristics
that might save your butt if you're lucky.