Re: khugepaged / firefox going wild in 3.18-rc

[Vlastimil Babka]

On 11/07/2014 02:57 PM, Norbert Preining wrote:
> Hi Vlastimil,
> 
> On Fri, 07 Nov 2014, Vlastimil Babka wrote:
>> Great, that's good news if I understand correctly, but ...
> 
> no "but ..."
> 
>> I suggested the commit to you for revert 1 day ago, and you say you
>> can't reproduce it for 2 days already? That's a bit suspicious. Did
> 
> No, you suggested it yesterday during the day, and here in Japan the
> next day is already over, so my feeling is two days ;-)
>
> So all fine ;-)

Ah, good, just wanted to be sure.
 
>> I'll prepare a debugging patch and send with instructions. Meanwhile
>> you could send the /proc/zoneinfo contents? :)
> 
> When? Running which kernel? Anyway, with the current kernel (reverted
> commit as before) I get the attached zoneinfo.

It doesn't matter which kernel. Thanks, but I didn't find anything
suspicious there... expected some oddly aligned zones, but it all seemed
to be aligned to pageblock boundaries.

> Thanks, and waiting for your patches ;-)

Tracing patch attached. You should apply this to the broken kernel, i.e.
without the revert, and have tracing enabled, i.e. CONFIG_FTRACE. There
should be /sys/kernel/debug/tracing directory.
To avoid overhead and noise, I would just run this kernel as usual, and
only when khugepaged/firefox/whatever starts misbehaving, do the following:

cd /sys/kernel/debug/tracing
echo 1 > tracing_on
echo 1 > events/compaction/enable
cat trace_pipe | tee /tmp/trace  #(or somewhere else)

You should see events scrolling on the screen, including the misbehaving
processes. After a minute or something, ctrl+c and send me the trace file.

Thanks a lot!
Vlastimil 


> Norbert
> 

------8<------