piping core dump to a program escapes container

[Shayan Pooya]

On Sun, Nov 1, 2015 at 12:25 PM, Richard Weinberger
<richard.weinberger at gmail.com> wrote:
> On Sat, Oct 24, 2015 at 11:54 PM, Shayan Pooya <shayan at liveve.org> wrote:
>> I noticed the following core_pattern behavior in my linux box while
>> running docker containers. I am not sure if it is bug, but it is
>> inconsistent and not documented.
>>
>> If the core_pattern is set on the host, the containers will observe
>> and use the pattern for dumping cores (there is no per cgroup
>> core_pattern). According to core(5) for setting core_pattern one can:
>>
>> 1. echo "/tmp/cores/core.%e.%p" > /proc/sys/kernel/core_pattern
>> 2. echo "|/bin/custom_core /tmp/cores/ %e %p " > /proc/sys/kernel/core_pattern
>>
>> The former pattern evaluates the /tmp/cores path in the container's
>> filesystem namespace. Which means, the host does not see a core file
>> in /tmp/cores.
>>
>> However, the latter evaluates the /bin/custom_core path in the global
>> filesystem namespace. Moreover, if /bin/core decides to write the core
>> to a path (/tmp/cores in this case as shown by the arg to
>> custom_core), the path will be evaluated in the global filesystem
>> namespace as well.
>>
>> The latter behaviour is counter-intuitive and error-prone as the
>> container can fill up the core-file directory which it does not have
>> direct access to (which means the core is also not accessible for
>> debugging if someone only has access to the container).
>>
>> Currently, I work around this issue by detecting that the process is
>> crashing from a container (by comparing the namespace pid to the
>> global pid) and refuse to dump the core if it is from a container.
>
> IMHO this is another example of the question whether a container should
> be a full blown Linux system or not.

I'd like to use containers as a way for separating concerns (IMHO this
is a popular use case for the containers, e.g. redhat's openshift).
The processes running inside a namespace should be that namespace's
problem.

> In your opinion containers should own core_pattern and be able to install
> their own dump helpers.

Well, it is not just my opinion. That's the default behaviour for when
you don't use a pipe in the core_pattern. The reason I'd say it is a
bug is because the two options are not consistent.

> Currently this is not the case as this is setting is only writable by
> the global root.
> Others might argue that containers are not a full virtual Linux with
> all features and
> an administrator may want to collect cores from all containers.

I agree there are some value in the admin getting a notification if
they choose to. The problem is that there is no clean mechanism for
the admin to send the core dump back to the namespace of the process.