On 05/31/2012 12:51 PM, Kamezawa Hiroyuki wrote:

quoted

One think to keep in mind: A file in memcg does not need to follow the
same format
of /proc/meminfo so we can bind mount. We should be able to
reconstruct that in
 userspace based on information available from the kernel. You can
even collect that
from multiple locations, and *then* you bind mount.

I'm sorry I couldn't fully understand. Could you explain more ?
Do you mean
  - bind mount memory cgroup directory into the container for exporting
information
  - Some user-space apps, FUSE-procfs or some, can provide enough
information

Implementation details aside, the idea is to have something like FUSE to 
hook the read(), and then construct the information it needs to present 
in the proper format.

Alternatively, for files that doesn't change a lot, you can create a 
file /container-storage-area/my_copy_of_meminfo at container creation, 
and bind mount *that* file.

There is no reason to bind mount a kernel-provided file directly.