[PATCH v2 3/3] Write dump into container's filesystem for pipe_type core_pattern

[Eric W. Biederman]

Mateusz Guzik <mguzik at redhat.com> writes:

> On Tue, Jun 07, 2016 at 07:29:37PM +0800, Zhao Lei wrote:
>> In current system, when we set core_pattern to a pipe, both pipe program
>> and program's output are in host's filesystem.
>> But when we set core_pattern to a file, the container will write dump
>> into container's filesystem.
>> 
>> For example, when we set following core_pattern:
>>  # echo "|/my_dump_pipe %s %c %p %u %g %t e" >/proc/sys/kernel/core_pattern
>> and trigger a segment fault in a container, my_dump_pipe is searched from
>> host's filesystem, and it will write coredump into host's filesystem too.
>> 
>> In a privileged container, user can destroy host system by following
>> command:
>>  # # In a container
>>  # echo "|/bin/dd of=/boot/vmlinuz" >/proc/sys/kernel/core_pattern
>>  # make_dump
>> 
>> Actually, all operation in a container should not change host's
>> environment, the container should use core_pattern as its private setting.
>> In detail, in core dump action:
>> 1: Search pipe program in container's fs namespace.
>> 2: Run pipe program in container's fs namespace to write coredump to it.
>> 
>> This patch fixed above problem by running pipe program with container's
>> fs_root.
>> 
>
> This does not look sufficient, but I can't easily verify.
>
> For instance, is the spawned process able to attach itself with ptrace
> to processes outside of the original container? Even if not, can it
> mount procfs and use /proc/pid/root of processes outside of said
> container?
>
> The spawned process should be subject to all limitations imposed on the
> container (which may mean it just must be in the container).

Pretty much.

The most constructive suggestion I have seen is to have the containers
init process logically fork in the kernel and spawn a process that way.

Anything that uses call_usermodehelper is trivially not safe because
of these concerns.

Eric