Re: ICE during kernel compile [Re: mipsel buildd kernels]
Goswin von Brederlow wrote:
> Thiemo Seufer <ica2_ts@csv.ica.uni-stuttgart.de> writes:
>
> > Goswin von Brederlow wrote:
> > > I have a script that methodically tries to remove lines from C source and
> > > recompiles it over and over to minimise such ICE. Removals that make
> > > the compile fail or fix the ICE (any output without the ICE) get
> > > undone, removals that still ICE are kept.
> >
> > That's a nice helper. :-) Trying that script on the preprocessed code
> > is probably the fastest method to find a good test case.
>
> Try that with a preprocessed C source that takes 3 1/2 hours to
> fail. Lets delete line 768-819. 3 1/2 hours later: Yep, still fails.
3.5 hours for twofish? Sounds like a CPU frequency of, err, 50 Hz. :-)
> But given enough time it does the job.
>
> > > But since I'm testbuilding the kernel-image debian package I haven't
> > > started that yet. The above problem could also just be an out of
> > > memory.
> >
> > Not catching an OOM condition is still a bug.
>
> How would you? malloc usually doesn't fail. Then at some point linux
> is all out of ram and something (not gcc) needs another page. Normaly
> lnux would kill 'something (not gcc)' but with the OOM killer option
> the 50MB using gcc gets killed instead, it being the one big mem eater.
Well, using memory overcommitment can be seen as feature.
IMHO it's a bug, as it won't work reliably.
> > > I only have 64MB and no swap so larger files can fail to build
> > > with no fault to gcc (usualy large C++ template riddled files but one
> > > never knows until one looks).
> >
> > Gcc/g++ _should_ stop with an appropriate error message in that case.
>
> It does when it runs out of virtual memory, but thats only when linux
> tells it a malloc failed. Seen that in the past. I'm not sure you can
> guard against the OOM killer.
Set /proc/sys/vm/overcommit_memory to 0, which is AFAIK the default.
Thiemo
Reply to: