Re: [atomic-devel] How to handle crashes

[Jakub Filak <jfilak redhat com>]

> On Thursday 12 of February 2015 10:53:50 Colin Walters wrote:
>

> On Tue, Feb 10, 2015, at 03:34 AM, Jakub Filak wrote:
>> The second idea is to put some core bits of ABRT directly to Atomic Host. My plan is to factor out ABRT's core_pattern helper (the thing we put to /proc/sys/kernel/core_pattern; man 5 core) that could be installed on Atomic host and create an image shipping the rest of ABRT stack. The footprint of such a design would be barely remarkable, but implementing it would ensure availability of debugging information in all cases.

> There was some work on having ABRT pick up crashes from systemd-coredump; is that still a work in progress?

First iteration of ABRT development has been finished, but I wouldn't suggest anybody use it in production (no one has tested it yet).

 

> Basically it would make sense to me for there to be an agent on the host to collect crashes; among other issues, an agent has to have visibility into which PID namespace/cgroup the core is coming from, and that's hard if it's already inside a container.

 

I thought that a super priviliged container sharing PID namespace with host would be able to get data from /host/proc/%P/, but I didn't try it, because /proc/sys/kernel/core_pattern cannot contain a path to a file in container, because kernel executes the pattern in host's namespace, so placing the core_pattern helper in a container is impossible.